Crowd-SAM: SAM as a Smart Annotator for

Object Detection in Crowded Scenes

Zhi Cai1,2 , Yingjie Gao1,2 , Yaoyan Zheng1,2

Nan Zhou1,2 , and Di Huang1,2⋆

SKLSDE, Beihang University, Beijing, China

1

IRIP Lab, SCSE, Beihang University, Beijing, China

2
arXiv:2407.11464v2 [cs.CV] 19 Jul 2024

{caizhi97, gaoyingjie, yaoyanzheng, zhounan0431, dhuang}@buaa.edu.cn

Abstract. Object detection is an important task that finds its applica-

tion in a wide range of scenarios. Generally, it requires extensive labels for
training, which is quite time-consuming, especially in crowded scenes. Re-
cently, Segment Anything Model (SAM) has emerged as a powerful zero-
shot segmenter, offering a novel approach to instance segmentation. How-
ever, the accuracy and efficiency of SAM and its variants are often com-
promised when handling objects in crowded scenes where occlusions often
appear. In this paper, we propose Crowd-SAM, a SAM-based framework
designed to enhance the performance of SAM in crowded scenes with
the cost of few learnable parameters and minimal labeled images. We
introduce an efficient prompt sampler (EPS) and a part-whole discrimi-
nation network (PWD-Net), facilitating mask selection and contributing
to an improvement in accuracy in crowded scenes. Despite its simplic-
ity, Crowd-SAM rivals state-of-the-art fully-supervised object detection
methods on several benchmarks including CrowdHuman and CityPer-
sons. Our code is available at https://github.com/FelixCaae/CrowdSAM.

Keywords: Detection in crowded scenes · Few-shot learning · Segment

anything model

1 Introduction
Object detection in crowded scenes is a fundamental task in areas such as au-
tonomous driving and video surveillance. The primary focus lies in identifying
and locating densely packed common objects like pedestrians and vehicles, where
occlusions present significant challenges. Great progress has been made in recent
years, including two-stage methods [45, 60] and query-based methods [8, 22, 62].
However, these methods mainly follow a supervised manner and necessitate ex-
tensive labeled training samples, incurring a considerable annotation cost of ap-
proximately 42.4 seconds per object [38] The density and complexity of crowded
scenes further aggravate the annotation burden. 1
⋆
Corresponding author
1
On average, an image contains approximately 22 objects in CrowdHuman [37] and
7 in MS COCO [24]
2 Z. Cai et al.

The high cost of collecting object annotations drives the exploration of alter-
natives such as few-shot learning [34, 39, 44], weakly supervised learning [41, 53],
semi-supervised learning [29,30,42,49], and unsupervised learning [1,6,21,26,48].
The best-performing ones, i.e. semi-supervised object detection (SSOD), lever-
age both labeled and unlabeled data for training and achieve a big success on
common benchmarks e.g. PASCAL VOC [15] and COCO [24]. Unfortunately,
SSOD introduces extra complexity such as complicated augmentations and on-
line pseudo-labeling.
Recently, prompt-based segmentation models have received increasing at-
tention due to their flexibility and scalability. Particularly, Segment Anything
Model (SAM) [20] show its high capability to effectively and accurately predict
the masks of regions specified by prompts, in any form of points, boxes, masks, or
text descriptions. Recognizing its exceptional potential, researchers have made
many efforts to adapt it for various vision tasks such as medical image recogni-
tion [31], remote sensing analysis [4, 12], industrial defect detection [52], etc.
Despite the huge progress [18,46,54] following SAM, applying SAM for object
detection in crowded scenes is seldom studied. In this paper, we investigate the
potential of SAM in such cases with two motivations. First, SAM is pre-trained
on a very large dataset i.e. SA-1B that contains most of the common objects
and it is thus reasonable to utilize the knowledge to facilitate labeling massive
data and training a brand-new detector. Second, SAM demonstrates a robust
segmentation ability in handling complicated scenes characterized by clustered
objects that are difficult for an object detector trained from scratch.
To this end, we propose Crowd-SAM, a smart annotator powered by SAM
for object detection in crowded scenes. As depicted in Fig. 1, we introduce a self-
promoting approach based on DINOv2 to alleviate the cost of human prompting.
Our method employs dense grids equipped with an Efficient Prompt Sampler
(EPS) to cover as many objects as possible at a moderate cost. To distinguish
the masks from multiple outputs precisely in occluded scenes, we design a mask
selection module, termed Part-Whole Discrimination Network (PWD-Net) that
learns to differentiate the output with the highest quality in Intersection over
Union (IoU) score. With a lightweight model design and fast training schedule,
it delivers considerable performance on public benchmarks including CrowdHu-
man [37] and CityPersons [58].
Our contributions can be summarized as follows:

1. We propose Crowd-SAM, a self-prompted segmentation method, for labeling

images containing clustered objects, producing accurate results with only a
few examples.
2. We design two novel components of Crowd-SAM, i.e. EPS and PWD-Net,
which effectively unleash the ability of SAM on crowded scenes.
3. We conduct comprehensive experiments on two benchmarks to demonstrate
the effectiveness and generalizable nature of Crowd-SAM.
 Crowd-SAM: Smart Annotation for Crowded Scenes 3

Self-Prompting &
Self-Selection
(a) SAM (b) Ours

Mask
SAM
Selection 标注图 Crowd-SAM 标注图
Adaptation
Annotated Results Annotated Results
Prompting Labeling
: Human Annotator

Fig. 1: Pipeline comparison between SAM and Crowd-SAM. Crowd-SAM only requires
a few labeled images and can automatically recognize target objects.

2 Related Work
Object Detection. General object detection aims to identify and locate objects
and is mainly divided into two categories: i.e. one-stage detectors and two-stage
detectors. One-stage detectors predict bounding boxes and class scores by using
image features [23, 27, 35], while two-stage detectors first generate region pro-
posals and then classify and refine them [9, 10, 36]. Recently, end-to-end object
detectors e.g. DETR [2, 55, 63] have replaced the hand-crafted modules such
as Non-Maximum Suppression (NMS) by adopting one-to-one matching in the
training phase, showing great potential in a wide variety of areas.
However, applying these detectors directly to pedestrian detection tends to
incur performance degradation due to the fact that pedestrians are often crowded
with occlusions appearing. Early work [32] proposes to integrate extra features
into a pedestrian detector to explore low-level visual cues, while follow-up meth-
ods [5, 56] attempt to utilize the head areas for better representation learning.
In [56], an anchor is associated with two targets, the whole body, and the head
part, to achieve a more robust detector from joint training. AdaptiveNMS [25]
adjusts the NMS threshold by predicting the density of pedestrians. Alternative
methods focus on the design of loss functions to improve the training process. For
example, RepLoss [45] encourages the prediction consistency of the same target
while repels the ones from different targets. Recently, Zheng et al . [62] models
the relation of queries to improve DETR-based detectors in crowded scenes and
achieves remarkable success. Although these works have pushed the boundaries
of object detection in crowded scenes to a new stage, they all rely on a large
number of labeled samples for training, which is labor-intensive. This limitation
inspires us to develop label-efficient detectors and automatic annotation tools,
with the help of SAM.
Few-Shot Object Detection (FSOD). This task aims to detect objects
of novel classes with limited training samples. FSOD methods can be roughly
classified into meta-learning based [16,51] and fine-tuning based ones [34,39,44].
Meta-RCNN [51] processes the query and support images in parallel via a siamese
network. The Region of Interest (RoI) features of the query are fused with the
class prototypes to effectively transfer knowledge learned from the support set.
TFA [44] proposes a simple two-stage fine-tuning method that only fine-tunes
the last layers of the network. FSCE [39] introduces a supervised contrastive loss
in the fine-tuning stage to mitigate misclassification issues. De-FRCN [34] stops
4 Z. Cai et al.

Crowd-SAM
: Tunable Module
DINO
Binary Classifier* : Frozen Module
Encoder

Foreground Mask
EPS

Input Image
SAM PWD-Net

Segmentation Results Detection Results

Part Whole Discrimination Net (PWD-Net) : Matrix Multiplication

DINO : Hardmard Product

Output
DINO Feature Binary Classifier*
+ : Element-wise Add
Softmax
C : Concatenate
Raw Mask Binarization

SAM Mask
Mask Token C MLP Selection
Output
Repeat
IoU Token IoU Head + IoU Scores

Fig. 2: The pipeline of Crowd-SAM shows the interaction between different modules.
DINO encoder and SAM are frozen in the training process. * represents the parameters
that are shared. For simplicity, the projection adapter of DINO is dismissed.

the gradient from the RPN and scales the gradient from R-CNN [36], followed
by a prototypical calibration block to refine the classification scores.
Segment Anything Models. SAM [20], a visual foundation model for seg-
mentation tasks, is trained on the SA-1B dataset using a semi-supervised learn-
ing paradigm. Its exposure to this vast repository of training samples renders it a
highly capable class-agnostic model, effectively handling a wide range of objects
in the world. Despite its impressive performance in solving segmentation tasks,
it suffers from several issues like domain shift, inefficiency, class-agnostic design,
etc. HQ-SAM [18] is proposed to improve its segmentation quality by learning a
lightweight adapter. Fast-SAM [50] and Mobile-SAM [54] focus on fastening the
inference speed of SAM by knowledge distillation. RSprompt [4] enables SAM to
generate semantically distinct segmentation results for remote sensing images by
generating appropriate prompts. Med-SA [47] presents a space-depth transpose
method to adapt 2D SAM to 3D medical images and a hyper-prompting adapter
to achieve prompt-conditioned adaptation. Unfortunately, these approaches ne-
cessitate a considerable amount of labeled data for effective adaptation, making
them impractical for crowded scenes where annotation costs are prohibitive. Dif-
ferent from them, Per-SAM [57] and Matcher [28] teach SAM to recognize spec-
ified objects with only one or few instances by extracting training-free similarity
priors. SAPNet [46] builds a weakly-supervised pipeline for instance segmenta-
tion. Although these approaches reduce data requirements, they still lag behind
the demands of crowded scenes, such as pedestrian detection, particularly with
occlusions.
 Crowd-SAM: Smart Annotation for Crowded Scenes 5

3 Method

3.1 Preliminaries

SAM [20] is a powerful and promising segmentation model that comprises three
main components: (a) an image encoder responsible for feature extraction; (b)
a prompt encoder designed to encode the geometric prompts provided by users;
and (c) a lightweight mask decoder that predicts the masks conditioned on the
given prompts. Leveraging extensive training data, SAM demonstrates impres-
sive zero-shot segmentation performance across various benchmarks. In particu-
lar, SAM makes use of points and boxes as prompts to specify interested regions.
DINO [3] represents a family of Vision Transformers (ViT) [7] learned in
a self-supervised manner designed for general-purpose applications. During its
training, DINO employs a self-distillation strategy akin to BYOL [11], fostering
the learning of robust representations. DINOv2 [33] strengthens the foundation
of DINO by integrating several additional pre-training tasks, improving its scala-
bility and stability, especially for large models e.g. ViT-H (1 billion parameters).
Thanks to its enhancement, DINOv2 shows a strong representation ability, in
particular for the task of semantic segmentation.

3.2 Problem Definition and Overall Framework

Problem Definition. As shown in Fig. 1, our goal is to detect objects (e.g.

pedestrians) in crowded scenes with few annotated data. We formulate this
problem as a one-class few-shot detection task. A common few-shot pipeline
is to divide data into the base split and the novel split. Differently, we directly
use the data of the target class for model training, as the foundation models
have already been trained on massive data. In particular, we employ segmenta-
tion masks as intermediate results which can be easily converted to bounding
boxes. During the training and evaluation processes, only box annotations are
provided.
Naive Study on SAM Auto-generator. The prompt number affects the
performance of SAM and we analyze this issue for crowded scenes. In this case,
we conduct several naive studies on CrowdHuman [37] with the auto generator of
SAM, which utilizes grid points to search every region. Tab. 1 conveys three key
observations: (1) dense grids are necessary for crowded scenes; (2) the ambiguity
of distributions of point prompts and class-agnostic prompts incur many false
positives (FPs); (3) the decoding time is a non-negligible burden when the grid
size is large. Therefore, the dense prompts and FP removal are key aspects
in designing SAM-based methods for detection/segmentation tasks in crowded
scenes.
Overall Framework. Inspired by the studies above, we equip SAM [20]
with several proper components to achieve an accurate and efficient annotation
framework, as illustrated in Fig. 2. To accurately locate clustered objects, we
employ the foundational model DINOv2 [33] to predict a semantic heat map,
a task that can be formulated with a simple binary classifier. To discriminate
6 Z. Cai et al.

Table 1: Comparison in terms of recall, average FPs, and decoding time (T ) of different
grid sizes (NG ) adopted by the SAM generator on CrowdHuman [37]. The oracle model
derives the prompts by computing the center of ground truth boxes. The decoding time
is collected on a 3090 Ti GPU card.

NG 16 32 64 128 Oracle
Recall 33.6 58.0 63.4 76.0 91.4
avg. FPs 51 112 227 485 -
T (s) 0.059 0.22 0.83 3.2 0.045

the output masks that is a mixture of correct masks, backgrounds, and part-
level masks, we design Part-Whole Discrimination Network (PWD-Net) that
takes as input both the learned tokens from SAM and the semantic-rich tokens
from DINOv2 to re-evaluate all the outputs. Finally, to handle the redundancy
brought by the use of dense grids, we propose an Efficient Prompt Sampler (EPS)
to decode the masks at a moderate cost.
We introduce the details of our methods in the following sections.

3.3 Class-specific Prompt Generation

Generating a unique point prompt for each object (e.g. pedestrian) in crowded
scenes is de facto a non-trivial problem. Thus, we take a step back and study
how to detect objects with multiple prompts associated with one object and
apply the proper post-processing techniques to remove duplicates. To this end,
we adopt a heatmap-based prompt generation pipeline that initially classifies
the regions and then generates prompts from the positive regions.
For the input image I ∈ RH×W ×3 , we first use a pre-trained image encoder
ED to extract semantic rich features. To better transfer the pre-trained features
to object segmentation, we add an MLP block after the final output layer and
H W
thus obtain the adapted features FDIN O ∈ R s × s ×C , where s is the patch size of
DINOv2 [33] and C is the output channel. Then, we employ a segmentation head
H W
HeadCLS to classify FDIN O pixel by pixel, resulting in a heatmap Ĥ ∈ [0, 1] s × s
that indicates the locations of objects as Ĥ = HeadCLS (FDIN O ).
Given only the bounding box annotations B ∈ [0, 1]NB ×4 , where NB is the
number of targets, this binary segmentation head can be optimized with box-
level supervision. However, the coarse boundaries tend to incur considerable
points scattering in background regions. To alleviate this issue, we use SAM [20]
to generate high-quality mask-level pseudo labels, which is illustrated in Fig. 4,
with ground truth (GT). The decoded masks are then merged into a single
foreground mask H ∈ {0, 1}256×256 . We use the dice loss for training the adapter
and segmentation head with the generated pseudo masks as follows:

{\cal L}_{fg} = dice(f(\hat {\mathbb {H}}), \mathbb {H}), \label {eq:bin_loss} (1)
where f is an up-sampling function that resizes H to 256×256. During inference,
we add a threshold t for mask binarization, which is simply set at 0.5 in our
 Crowd-SAM: Smart Annotation for Crowded Scenes 7

experiments. The binarized masks are mapped to point prompts PG which only
contain those in positive regions.

3.4 Semantic-guided Mask Prediction

Given the proposals generated in Sec. 3.3, our further aim is to efficiently decode
the dense prompts and accurately discriminate the generated masks. As depicted
in Fig. 3, each instance contains a set of prompts due to the density of grids.
Supposing that only one in-position prompt is required for mask prediction in
SAM, decoding all the prompts would lead to not only a waste of computation
but also more FPs for some poorly located prompts.

Efficient Prompt Sampler (EPS) : Pruned Prompt : Selected Prompt

SAM
Iteration 0:
PWD-Net
X X ...
... Valid Mask
Iteration n: X X X Inputs Results

Fig. 3: Illustration of EPS. PWD-Net produces valid masks with a threshold. In each
iteration, we prune prompts (with a cross above) that fall inside valid masks .

Efficient Prompt Sampler (EPS). To address this challenge, we intro-

duce EPS, which dynamically prunes prompts based on the confidence of decoded
masks. This method is elaborated in Algorithm 1. Beginning with the generated
point prompts list PG , EPS, in each iteration, samples a batch of prompts PB
from PG using uniform random sampling. The sampled prompts PB are then ap-
pended to the output points list PS . Subsequently, we employ the SAM generator
with the batched prompts PB to produce masks M . Furthermore, we aggregate
discriminant confidence scores S from PWD-Net, which are employed to select
valid masks M ′ using a score threshold T , by using M ′ = M [S > T ]. Refer to
a detailed exposition of PWD-Net in the subsequent section. The valid masks
represent regions that are believed to be well segmented, and we thus remove the
points that have already been covered by any mask m in M ′ from the prompts
list PG . The iteration halts when PG is empty. Additionally, EPS establishes a
stopping criterion with a parameter K, terminating the sampling process once
the total sample count reaches this limit. This parameter is instrumental in man-
aging the overall decoding cost. In our experiments, K is empirically set to 500
to strike a balance between efficiency and recall.
Part-Whole Discrimination Network (PWD-Net). Given the raw masks
predicted by SAM conditioned on the sampled batch of prompts PB , we design
an automatic selector for choosing the best-fitting mask. It is expected to opti-
mize the outputs in two aspects: (i) refining the output IoU score according to
8 Z. Cai et al.

Algorithm 1: Process of Efficient Prompt Sampler (EPS).

Input: feature extracted from image I by image encoder of SAM: FSAM ;
generated prompt list: PG = {p1 , p2 , ..., pN }; SAM generator: G; mask
confidence threshold: T ; target sampled prompt list size: K
Output: sampled prompt list: PS ; valid mask list: MS
1 PS ← ∅;
2 while |PG | > 0 and |PS | < K do
3 Sample a batch of prompts PB ⊆ PG with uniform random sampling;
4 PS ← PS ∪ PB ;
5 PG ← PG \ PB ;
6 Generate masks M corresponding to PB by G(FSAM , PB );
7 Select valid masks M ′ according to T from M ;
8 MS ← MS ∪ M ′ ;
9 for p ∈ PG do
10 if ∃m ∈ M ′ such that p ∈ m then
11 PG ← PG \ {p}
12 end
13 end
14 end

the quality of related masks if they are positive samples and (ii) suppressing the
scores of samples that fall in background regions.
Illustrated in Fig. 2, for the masks generated corresponding to N prompts,
we leverage the Mask Tokens and IoU Tokens within the mask decoder of SAM
along with the sophisticated features extracted by the self-supervised pre-trained
model DINOv2 [33]. M and U are responsible for mask decoding and IoU pre-
diction in the SAM Mask Decoder, respectively. Thus, we suppose that they
contain shape-aware information, which is helpful in discriminating the mask.
These components enable us to compute a discriminant confidence score S for
each specific prompt in a few-shot scenario. Initially, the refined IoU score Siou
is computed as follows:

S_{iou}=\text {Head}_{par}(\text {Concat}(\text {Repeat}({\cal U}), {\cal M}))+\text {Head}_{IoU}({\cal U}), \label {eq:s_iou} (2)

where Headpar is a parallel IoU head, consisting of an MLP block; U ∈ RN ×1×C

and M ∈ RN ×4×C denote the IoU Tokens and Mask Tokens respectively. No-
tice in Eq. (2), Siou is a sum of outputs from two individual heads, the parallel
adapter head Headpar and the original HeadIoU . We freeze the parameters of
HeadIoU to avoid overfitting in few-shot learning. Moreover, since these two to-
kens, M and U have different shapes, we repeat U by 4 times and concatenate
these two tokens in the channel dimension. The refined score Siou ∈ RN ×4 en-
capsulates the quality of the generated masks by assessing the texture-aware
feature from M and U.
Further, by harnessing the semantic feature embedded in the self-supervised
pre-trained model DINOv2, along with the mask data M̂ , we calculate the dis-
 Crowd-SAM: Smart Annotation for Crowded Scenes 9

criminant score Scls :

S_{cls} = \sigma ({\text {Head}_{CLS}({\cal O})}), {\cal O} = \text {Pool}(d(\text {Softmax}(\hat {M})) \circ F_{DINO}). (3)

Here, O ∈ RN ×C denotes the extracted Semantic Token, and M̂ ∈ RN ×H×W ×4

and FDIN O ∈ RN ×h×w×C represent the masks generated by SAM [31] and
the features extracted by DINOv2, respectively. We denote d as a down-scale
function that resizes the mask to h × w, consistent with FDIN O . Pool is a global
pooling function that conducts mean pooling on the x-axis and y-axis and ◦ is the
Hadamard product. The discriminant score Scls ∈ RN ×4 predicts whether masks
belong to the foreground or background. HeadCLS shares the same parameters
with the binary classifier introduced in Sec. 3.3. Finally, we calculate a joint score
of discrimination and estimate the quality for masks by simply multiplying the
two scores: S = Siou · Scls .
During training, prompts sampled from real masks are taken as input. Re-
garding the prompts within the foreground, the discriminant confidence score
aims to accurately predict the IoU of the generated and real masks. Conversely,
for the prompts within the background, the score is ideally 0. Hence, the loss
function for this aspect is formulated as follows:

s^i_{target} = \begin {cases} IoU(m^i,m^i_{GT}), & m^i_{GT} \in M^{bg}_{GT}, \\ 0, & m^i_{GT} \in M^{fg}_{GT} , \end {cases} (4)

{\cal L}_{iou}=\text {MSE}(S,S_{target})&.\label {eq:iou_loss} (5)

Here, sitarget denotes the target score of the mask mi generated by the ith prompt,
bg fg
and Starget = {sitarget }N i N i N
i=1 ; M = {m }i=1 ; MGT = {mGT }i=1 = MGT ∪ MGT .

3.5 Training and Inference

The total training loss of the entire framework combines Eq. (1) and Eq. (5):
{\cal L} = {\cal L}_{fg} + {\cal L}_{iou} (6)
At inference, we select the mask with the highest confidence score S among
the 4 masks as the output of PWD-Net, which we denote as M ∈ RN ×H×W and
So ∈ RN . We adopt a window cropping strategy to enhance the performance on
small objects as [20]. This strategy slices the whole image into overlapping crops
where each crop is individually processed. The final results are merged from the
outputs of each crop and here we apply NMS to remove duplicate proposals.

4 Experiments
Datasets. Following [62], we adopt CrowdHuman [37] as the benchmark to con-
duct main experiments and ablation studies. CrowdHuman [37] collects and an-
notates images containing crowded persons from the Internet. It contains 15,000,
10 Z. Cai et al.

Table 2: Comparative results (%) on CrowdHuman [37] val. All the SAM-based meth-
ods adopt ViT-L [7] as the pre-trained backbone (SRCNN denotes Sparse R-CNN [40],
a baseline in [62] and * represents using the multi-crop trick).

Method Backbone #Shot AP MR−2 Recall Secs/Img

Fully supervised object detectors
ATSS [59] ResNet-50 [14] Full 80.3 59.7 86.1 0.051
FCOS [43] ResNet-50 [14] Full 76.3 65.5 82.6 0.045
Iter-SRCNN [62] ResNet-50 [14] Full 85.9 58.3 93.3 0.25
DINO [55] ResNet-50 [14] Full 86.7 57.6 94.5 0.27
Few-shot object detectors
TFA [44] ResNet-101 [14] 10 46.9 84.3 57.9 0.067
FSCE [39] ResNet-101 [14] 10 43.0 84.7 50.0 0.072
De-FRCN [34] ResNet-101 [14] 10 46.4 85.9 65.5 0.072
SAM-based approaches
SAM [20] ViT-L [7] 0 - - 65.6 1.3
SAM* [20] ViT-L [7] 0 - - 79.6 6.7
Matcher [34] ViT-L [7] 1 8.0 88.9 23.9 22.0
Crowd-SAM ViT-L [7] 10 71.4 75.1 83.9 1.7
Crowd-SAM* ViT-L [7] 10 78.4 74.8 85.6 8.1

4,370, and 5,000 images for training, validation, and testing, respectively. We also
evaluate our method on CityPersons [58] for a realistic urban-scene scenario. Ad-
ditionally, we utilize OCC-Human [61], which is specially reputed for occluded
persons. For these pedestrian datasets, we use visible annotations (only including
visible areas of an object) for training and evaluation. To validate the extensi-
bility of Crowd-SAM, we further devise a multi-class version of Crowd-SAM by
adding a multi-class classifier. We employ 0.1 % percent of the COCO [24] train-
val set for training and the COCO val set for validation. Besides, we validate
our method on a subset with occluded objects on COCO, i.e. COCO-OCC [17],
extracted by selecting the images whose objects have a high overlapping ratio.
Implementation Details. We utilize SAM (ViT-L) [20] and DINOv2 (ViT-
L) [33] as the base models for all experiments. In the fine-tuning stage, all their
parameters are frozen to avoid over-fitting. Instead of real GT, we use the pseudo
masks generated by SAM as Starget to supervise the learning of PWD-Net in
Eq. (5). These generated pseudo labels are of high quality as shown in Fig. 4(c).
We randomly pick the points from the pseudo masks as positive training samples
and the ones from the background as negative training samples. In the training
process, we use Adam [19] with a learning rate of 10−5 , a weight decay of 10−4 ,
β1 = 0.9, and β2 = 0.99 for optimization. We train our module for 2,000 itera-
tions with a batch size of 1, which can be done on a single GTX 3090 Ti GPU
in several minutes. For more details, please refer to the Appendix.
Evaluation Metrics. Following [62], we use AP with IoU threshold at 0.5,
MR−2 , and Recall as metrics. Generally, a higher AP, Recall, and lower MR−2
value indicates better performance.
 Crowd-SAM: Smart Annotation for Crowded Scenes 11

4.1 Experimental Results on Pedestrian Detection

For fair comparison, we re-implement the counterparts [34,43,55,59,62] with a 2×

schedule using visible annotations in the CrowdHuman [37] and CityPersons [58]
datasets.
Main Results. We compare Crowd-SAM with most related methods includ-
ing fully-supervised object detectors [43, 55, 59, 62], few shot object detectors [34]
and SAM-based methods [20,28]. Notice that we use visible annotations for which
derive different results from those on full box annotations.
As Tab. 2 shows, with only 10 labeled images, Crowd-SAM achieves com-
parable performance with full-supervised object detectors whose best result is
86.7% AP, delivered by DINO [55]. Particularly, Crowd-SAM outperforms an
advanced anchor-free detector FCOS [43] by 2.1% AP. These results indicate
that by using proper adaptation techniques, SAM can reach very competitive
performance on complex pedestrian detection datasets like CrowdHuman. On
the other side, Crowd-SAM achieves SOTA performance on few-shot detection
settings and outperforms all the few-shot object detectors by a significant mar-
gin. DeFRCN [34] is a well-established few-shot object detector equipped with
ResNet-101 [14] that reports 46.4% in AP. Notably, our method exceeds it by
32%, indicating the superiority in the few-shot detection setting. The qualitative
comparison between Crowd-SAM and De-FRCN is depicted in Fig. 4. For SAM-
based methods, our Crowd-SAM largely leads the SAM baseline by 6% (with
multi-crop) and 18.3% (w/o multi-crop) in Recall. Besides, Crowd-SAM is also
superior to the other methods such as Matcher [28].
Results on OccHuman and CityPersons. To investigate the performance
of Crowd-SAM in occluded scenes, we compare it with an advanced counter-
part Pose2Seg [61], and the results are shown in Tab. 3. It is noteworthy that
Pose2Seg is a fully-supervised detector while Crowd-SAM is a few-shot detector.
As can be seen from the results, Crowd-SAM leads Pose2Seg by 9.2% AP which
demonstrates its robustness in occluded scenes. Also, we apply our method to
a not-so-crowded but more realistic urban dataset, i.e. CityPersons [58], as re-
ported in Tab. 4. Crowd-SAM outperforms TFA [44] by 2.7% AP, FSCE [39] by
1.1% AP, and De-FRCN [34] by 7.8% AP. In conclusion, our method remains
competitive compared to advanced few-shot object detectors, even though it is
not specifically designed for sparse scenes.
These results illustrate that Crowd-SAM well unleashes the power of vision
foundation models, i.e. SAM and DINO, in all of the crowded, occluded, and
urban scenes.

4.2 Experimental Results on Multi-class Object Detection

To further explore the extensibility in a more popular setting, we devise a multi-

class Crowd-SAM. The multi-class version of Crowd-SAM is slightly different by
replacing the binary classifier with a multi-class one. We then validate multi-
class Crowd-SAM on COCO [24], a widely adopted common object detection
12 Z. Cai et al.

Table 3: Comparative results (%) on OccHuman val, where APM and APH represent
AP in moderate and hard cases according to occlusion ratios, respectively.

Method Backbone AP APM APH

Mask R-CNN [13] ResNet50-FPN [14] 16.3 19.4 11.3
Pose2Seg [61] ResNet50-FPN [14] 22.2 26.1 15.0
Crowd-SAM ViT-L [7] 31.4 26.5 17.7

Table 4: Comparative results (%) on CityPersons val.

Method # Shot AP MR−2

FCOS [43] Full 58.8 30.0
ATSS [59] Full 54.1 27.8
Iter-SRCNN [62] Full 57.9 31.0
TFA [44] 50 30.6 53.8
FSCE [39] 50 32.2 46.5
De-FRCN [34] 50 25.5 67.1
Crowd-SAM 50 33.3 31.7

benchmark, and COCO-OCC [17] which is a split of COCO that mainly consists
of images with a high occlusion ratio.
We compare our method with two supervised detectors, i.e. Faster R-CNN [36]
and BCNet [17], and report the results in Tab. 5. It can be seen that our Crowd-
SAM is comparable to the supervised detectors on both datasets and drops only
1.4 AP% when comparing those of COCO and COCO-OCC. This minor drop
indicates that our method is robust to occlusions.

4.3 Ablation Studies

Ablation on Modules. We conduct ablation studies on the key components
of Crowd-SAM, including foreground location, EPS, and PWD-Net, to validate
their effectiveness. In Tab. 6, the performance in AP significantly drops by 7.4%
and recall by 8.5% when FG location is removed, indicating the importance of
restricting foreground areas. As for EPS, once it is removed, the AP drops by
0.6%. We suppose that EPS not only accelerates the sampling process of dense
prompts but also helps focus on the difficult part of the image, which conveys
ambiguous semantics. We also compare EPS with some other batch iterators
in Tab. 7. Besides, we find that PWD-Net is indispensable and when removed,
the AP dramatically drops to 17.0%. Finally, we point out that multi-cropping is
a strong trick to enhance the performance which contributes 7.0% AP to the final
performance. Overall, these results prove that all the components are essential.
Ablation on EPS. We conduct a comprehensive study on EPS by compar-
ing it with several variants. We forward each image only once to avoid the extra
latency caused by multi-crop. We use a default sampler that iterates through all
training samples and a random iterator with a halting threshold K as counter-
parts. As reported in Tab. 7, the AP and Recall increase with the grid size for
 Crowd-SAM: Smart Annotation for Crowded Scenes 13

Table 5: Comparative results (%) on COCO val and COCO-OCC.

Methods Backbone COCO-OCC COCO

AP AP50 AP AP50
Faster R-CNN [36] ResNet50-FPN [14] 29.7 50.0 33.5 53.7
BCNet [17] ResNet50-FPN [14] 31.7 51.1 34.6 54.4
Crowd-SAM ViT-L [7] 20.6 31.5 22.0 33.7

Table 6: Ablation results (%) on the main components in Crowd-SAM. FG location

represents the use of the binary classifier on DINOv2 features. The last line represents
a SAM baseline with a standard 32 × 32 grid as inputs.

FG loc. EPS PWD-Net Multi-Crop AP MR−2 Recall

✓ ✓ ✓ ✓ 78.4 74.8 85.6
✓ ✓ ✓ 71.0 77.9 77.1
✓ ✓ ✓ 77.8 71.8 83.9
✓ ✓ ✓ 17.0 99.6 83.1
✓ ✓ ✓ 71.4 75.1 83.9
✓ 8.7 99.8 70.1

all three samplers. However, the default sampler suffers an out-of-memory error
when the grid size reaches 128, preventing it from being adopted in this setting.
As for the random sampler, its performance is constrained by K and a grid size
larger than 64 only leads to limited improvement, e.g. 0.1% AP. On the contrary,
our EPS benefits from a much larger grid size and achieves a better result.
Ablation on PWD-Net. We compare PWD-Net to the variants that re-
place some tokens with a full-zero placeholder. We also consider two designs, di-
rectly tuning the IoU head or learning a parallel IoU head. As shown in Tab. 8,
all the three tokens, i.e. Mask Token M, IoU Token U, and Semantic Token O,
contribute to the final result. Particularly, once M is removed, the AP drops by
40.0%, which is a catastrophic decline. This degradation suggests that the mask
token contains the shape-aware feature that is essential for the part-whole dis-
crimination task. Notably, the AP drops by 2.8% when we tune the pre-trained
IoU Head, suggesting that it is prone to overfit the few labeled images. By freez-

Table 7: Comparison (%) of different samplers on CrowdHuman [37] val. Full means
using all prompts. OOM represents out-of-memory errors which occur when the GPU
memory is all consumed.

Full Random EPS

Grid K AP Recall AP Recall AP Recall
32 × 32 500 57.6 60.5 57.6 60.5 57.0 60.0
64 × 64 500 69.4 73.9 69.4 73.9 69.3 73.6
128 × 128 500 OOM 69.7 74.1 72.4 77.8
192 × 192 500 OOM 69.8 73.7 73.2 78.2
256 × 256 500 OOM 68.9 73.0 72.3 78.0
14 Z. Cai et al.

Table 8: Ablation results (%) on the design of PWD-Net. M, U, and O represent the
mask token, IoU token, and semantic token, respectively. For the IoU head, F means
freezing the original IoU head and training a parallel one, and T indicates tuning the
original IoU head.

M U O IoU head AP MR−2

✓ ✓ ✓ F 78.4 74.8
✓ ✓ ✓ T 75.6(-2.8) 80.4(+5.6)
✓ ✓ F 77.3(-1.1) 73.5(-1.3)
✓ ✓ F 76.3(-2.1) 74.8(-0.0)
✓ ✓ F 38.4(-40.0) 95.8(+21.0)

ing the IoU head of SAM, PWD-Net can benefit more from the shape-aware
knowledge learned from massive segmentation data.

Fig. 4: Qualitative comparison between Crowd-SAM (a) and De-FRCN (b). Crowd-
SAM predictions are more accurate especially in the boundaries of persons. We also
plot the GT boxes (blue rectangles) and the generated masks (yellow regions), which
are of high quality (c). In (d ), we plot our prompt filtering results, where preserved
prompts (red points) are much fewer than the removed ones (gray points). Zoom in for
a better view.

5 Conclusion
This paper proposes Crowd-SAM, a SAM-based framework, for object detection
and segmentation in crowded scenes, designed to streamline the annotation pro-
cess. For each image, Crowd-SAM generates dense prompts for high recall and
uses EPS to prune redundant prompts. To achieve accurate detection in occlu-
sion cases, Crowd-SAM employs PWD-Net which leverages several informative
tokens to select the masks that best fit. Combined with the proposed modules,
Crowd-SAM achieves 78.4% AP on CrowdHuman, comparable to full-supervised
detectors, validating that object detection in crowded scenes can benefit from
foundation models like SAM with data efficiency.
 Crowd-SAM: Smart Annotation for Crowded Scenes 15

Acknowledgements

This work is partly supported by the National Natural Science Foundation of

China (No. 62022011), the Research Program of State Key Laboratory of Soft-
ware Development Environment, and the Fundamental Research Funds for the
Central Universities.

References

1. Bar, A., Wang, X., Kantorov, V., Reed, C.J., Herzig, R., Chechik, G., Rohrbach,
A., Darrell, T., Globerson, A.: Detreg: Unsupervised pretraining with region priors
for object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 14605–14615
(2022)
2. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-
to-end object detection with transformers. In: Eur. Conf. Comput. Vis. pp. 213–
229. Springer (2020)
3. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.:
Emerging properties in self-supervised vision transformers. In: Int. Conf. Comput.
Vis. pp. 9650–9660 (2021)
4. Chen, K., Liu, C., Chen, H., Zhang, H., Li, W., Zou, Z., Shi, Z.: Rsprompter: Learn-
ing to prompt for remote sensing instance segmentation based on visual foundation
model. IEEE Transactions on Geoscience and Remote Sensing (2024)
5. Chi, C., Zhang, S., Xing, J., Lei, Z., Li, S.Z., Zou, X.: Pedhunter: Occlusion robust
pedestrian detector in crowded scenes. In: AAAI. vol. 34, pp. 10639–10646 (2020)
6. Dai, Z., Cai, B., Lin, Y., Chen, J.: Up-detr: Unsupervised pre-training for object
detection with transformers. In: IEEE Conf. Comput. Vis. Pattern Recog. pp.
1601–1610 (2021)
7. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner,
T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth
16x16 words: Transformers for image recognition at scale. In: Int. Conf. Learn.
Represent. (2021)
8. Gao, F., Leng, J., Gan, J., Gao, X.: Selecting learnable training samples is all
detrs need in crowded pedestrian detection. In: ACM Int. Conf. Multimedia. pp.
2714–2722 (2023)
9. Girshick, R.: Fast r-cnn. In: Int. Conf. Comput. Vis. pp. 1440–1448 (2015)
10. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for ac-
curate object detection and semantic segmentation. In: IEEE Conf. Comput. Vis.
Pattern Recog. pp. 580–587 (2014)
11. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doer-
sch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own
latent-a new approach to self-supervised learning. Adv. Neural Inform. Process.
Syst. 33, 21271–21284 (2020)
12. Gui, S., Song, S., Qin, R., Tang, Y.: Remote sensing object detection in the deep
learning era—a review. Remote Sensing 16(2), 327 (2024)
13. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: Int. Conf. Comput.
Vis. pp. 2961–2969 (2017)
14. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition.
In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 770–778 (2016)
16 Z. Cai et al.

15. Hoiem, D., Divvala, S.K., Hays, J.H.: Pascal voc 2008 challenge. World Literature
Today 24(1), 1–4 (2009)
16. Kang, B., Liu, Z., Wang, X., Yu, F., Feng, J., Darrell, T.: Few-shot object detection
via feature reweighting. In: Int. Conf. Comput. Vis. pp. 8420–8429 (2019)
17. Ke, L., Tai, Y.W., Tang, C.K.: Deep occlusion-aware instance segmentation with
overlapping bilayers. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4019–4028
(2021)
18. Ke, L., Ye, M., Danelljan, M., Tai, Y.W., Tang, C.K., Yu, F., et al.: Segment
anything in high quality. In: Adv. Neural Inform. Process. Syst. vol. 36 (2024)
19. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: Int. Conf.
Learn. Represent. (2015)
20. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T.,
Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Int. Conf.
Comput. Vis. pp. 4015–4026 (2023)
21. Li, M., Wu, J., Wang, X., Chen, C., Qin, J., Xiao, X., Wang, R., Zheng, M., Pan,
X.: Aligndet: Aligning pre-training and fine-tuning in object detection. In: Int.
Conf. Comput. Vis. pp. 6866–6876 (2023)
22. Lin, M., Li, C., Bu, X., Sun, M., Lin, C., Yan, J., Ouyang, W., Deng, Z.: Detr for
crowd pedestrian detection. arXiv preprint arXiv:2012.06785 (2020)
23. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollár, P.: Focal loss for dense object
detection. In: Int. Conf. Comput. Vis. pp. 2980–2988 (2017)
24. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P.,
Zitnick, C.L.: Microsoft coco: Common objects in context. In: Eur. Conf. Comput.
Vis. pp. 2906–2917 (2014)
25. Liu, S., Huang, D., Wang, Y.: Adaptive nms: Refining pedestrian detection in a
crowd. In: IEEE Conf. Comput. Vis. Pattern Recog. (June 2019)
26. Liu, S., Li, Z., Sun, J.: Self-emd: Self-supervised object detection without imagenet.
arXiv preprint arXiv:2011.13677 (2020)
27. Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C.Y., Berg, A.C.: Ssd:
Single shot multibox detector. In: Eur. Conf. Comput. Vis. pp. 21–37. Springer
(2016)
28. Liu, Y., Zhu, M., Li, H., Chen, H., Wang, X., Shen, C.: Matcher: Segment anything
with one shot using all-purpose feature matching. In: Int. Conf. Learn. Represent.
(2024)
29. Liu, Y.C., Ma, C.Y., He, Z., Kuo, C.W., Chen, K., Zhang, P., Wu, B., Kira, Z.,
Vajda, P.: Unbiased teacher for semi-supervised object detection. In: Int. Conf.
Learn. Represent. (2021)
30. Liu, Y.C., Ma, C.Y., Kira, Z.: Unbiased teacher v2: Semi-supervised object de-
tection for anchor-free and anchor-based detectors. In: IEEE Conf. Comput. Vis.
Pattern Recog. pp. 9819–9828 (2022)
31. Ma, J., He, Y., Li, F., Han, L., You, C., Wang, B.: Segment anything in medical
images. Nature Communications 15(1), 654 (2024)
32. Mao, J., Xiao, T., Jiang, Y., Cao, Z.: What can help pedestrian detection? In:
IEEE Conf. Comput. Vis. Pattern Recog. pp. 3127–3136 (2017)
33. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V.,
Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust
visual features without supervision. Trans. Mach. Learn Res. (2024)
34. Qiao, L., Zhao, Y., Li, Z., Qiu, X., Wu, J., Zhang, C.: Defrcn: Decoupled faster
r-cnn for few-shot object detection. In: Int. Conf. Comput. Vis. pp. 8681–8690
(2021)
 Crowd-SAM: Smart Annotation for Crowded Scenes 17

35. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified,
real-time object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 779–
788 (2016)
36. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detec-
tion with region proposal networks. In: Adv. Neural Inform. Process. Syst. vol. 28
(2015)
37. Shao, S., Zhao, Z., Li, B., Xiao, T., Yu, G., Zhang, X., Sun, J.: Crowdhuman:
A benchmark for detecting human in a crowd. arXiv preprint arXiv:1805.00123
(2018)
38. Su, H., Deng, J., Fei-Fei, L.: Crowdsourcing annotations for visual object detection.
In: Workshops at the twenty-sixth AAAI conference on artificial intelligence (2012)
39. Sun, B., Li, B., Cai, S., Yuan, Y., Zhang, C.: Fsce: Few-shot object detection via
contrastive proposal encoding. In: IEEE Conf. Comput. Vis. Pattern Recog. pp.
7352–7362 (2021)
40. Sun, P., Zhang, R., Jiang, Y., Kong, T., Xu, C., Zhan, W., Tomizuka, M., Li, L.,
Yuan, Z., Wang, C., et al.: Sparse r-cnn: End-to-end object detection with learnable
proposals. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 14454–14463 (2021)
41. Tang, P., Wang, X., Bai, X., Liu, W.: Multiple instance detection network with
online instance classifier refinement. In: IEEE Conf. Comput. Vis. Pattern Recog.
pp. 2843–2851 (2017)
42. Tang, Y., Chen, W., Luo, Y., Zhang, Y.: Humble teachers teach better students
for semi-supervised object detection. In: IEEE Conf. Comput. Vis. Pattern Recog.
pp. 3132–3141 (2021)
43. Tian, Z., Shen, C., Chen, H., He, T.: Fcos: Fully convolutional one-stage object
detection. In: Int. Conf. Comput. Vis. pp. 9627–9636 (2019)
44. Wang, X., Huang, T.E., Darrell, T., Gonzalez, J.E., Yu, F.: Frustratingly simple
few-shot object detection. Int. Conf. Mach. Learn. (2020)
45. Wang, X., Xiao, T., Jiang, Y., Shao, S., Sun, J., Shen, C.: Repulsion loss: Detecting
pedestrians in a crowd. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7774–7783
(2018)
46. Wei, Z., Chen, P., Yu, X., Li, G., Jiao, J., Han, Z.: Semantic-aware sam for point-
prompted instance segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp.
3585–3594 (2024)
47. Wu, J., Fu, R., Fang, H., Liu, Y., Wang, Z., Xu, Y., Jin, Y., Arbel, T.: Medical
sam adapter: Adapting segment anything model for medical image segmentation.
arXiv preprint arXiv:2304.12620 (2023)
48. Xie, E., Ding, J., Wang, W., Zhan, X., Xu, H., Sun, P., Li, Z., Luo, P.: Detco:
Unsupervised contrastive learning for object detection. In: Int. Conf. Comput. Vis.
pp. 8392–8401 (2021)
49. Xu, M., Zhang, Z., Hu, H., Wang, J., Wang, L., Wei, F., Bai, X., Liu, Z.: End-
to-end semi-supervised object detection with soft teacher. In: Int. Conf. Comput.
Vis. pp. 3060–3069 (2021)
50. Xu, Z., Wenchao, D., Yongqi, A., Yinglong, D., Tao, Y., Min, L., Ming, T., Jinqiao,
W.: Fast segment anything. arXiv preprint arXiv:2306.12156 (2023)
51. Yan, X., Chen, Z., Xu, A., Wang, X., Liang, X., Lin, L.: Meta r-cnn: Towards
general solver for instance-level low-shot learning. In: Int. Conf. Comput. Vis. pp.
9577–9586 (2019)
52. Ye, Z., Lovell, L., Faramarzi, A., Ninic, J.: Sam-based instance segmentation mod-
els for the automation of masonry crack detection. arXiv preprint arXiv:2401.15266
(2024)
18 Z. Cai et al.

53. Zeng, Z., Liu, B., Fu, J., Chao, H., Zhang, L.: Wsod2: Learning bottom-up and
top-down objectness distillation for weakly-supervised object detection. In: Int.
Conf. Comput. Vis. (2019)
54. Zhang, C., Han, D., Qiao, Y., Kim, J.U., Bae, S.H., Lee, S., Hong, C.S.: Faster
segment anything: Towards lightweight sam for mobile applications. arXiv preprint
arXiv:2306.14289 (2023)
55. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L., Shum, H.Y.: DINO:
DETR with improved denoising anchor boxes for end-to-end object detection. In:
Int. Conf. Learn. Represent. (2023)
56. Zhang, K., Xiong, F., Sun, P., Hu, L., Li, B., Yu, G.: Double anchor r-cnn for
human detection in a crowd. arXiv preprint arXiv:1909.09998 (2019)
57. Zhang, R., Jiang, Z., Guo, Z., Yan, S., Pan, J., Dong, H., Qiao, Y., Gao, P., Li, H.:
Personalize segment anything model with one shot. In: Int. Conf. Learn. Represent.
(2024)
58. Zhang, S., Benenson, R., Schiele, B.: Citypersons: A diverse dataset for pedestrian
detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3213–3221 (2017)
59. Zhang, S., Chi, C., Yao, Y., Lei, Z., Li, S.Z.: Bridging the gap between anchor-
based and anchor-free detection via adaptive training sample selection. In: IEEE
Conf. Comput. Vis. Pattern Recog. pp. 9759–9768 (2020)
60. Zhang, S., Wen, L., Bian, X., Lei, Z., Li, S.Z.: Occlusion-aware r-cnn: Detecting
pedestrians in a crowd. In: Eur. Conf. Comput. Vis. pp. 637–653 (2018)
61. Zhang, S.H., Li, R., Dong, X., Rosin, P., Cai, Z., Han, X., Yang, D., Huang, H.,
Hu, S.M.: Pose2seg: Detection free human instance segmentation. In: IEEE Conf.
Comput. Vis. Pattern Recog. pp. 889–898 (2019)
62. Zheng, A., Zhang, Y., Zhang, X., Qi, X., Sun, J.: Progressive end-to-end object
detection in crowded scenes. In: IEEE Conf. Comput. Vis. Pattern Recog. pp.
857–866 (2022)
63. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable
transformers for end-to-end object detection. In: Int. Conf. Learn. Represent.
(2021)