Object Detection With Probabilistic Guarantees

Florence de Grancey, Jean-Luc Adam, Lucian Alecu, Sébastien Gerchinovitz,

Franck Mamalet, David Vigouroux

To cite this version:

Florence de Grancey, Jean-Luc Adam, Lucian Alecu, Sébastien Gerchinovitz, Franck Mamalet, et
al.. Object Detection With Probabilistic Guarantees: a Conformal Prediction Approach. Fifth Inter-
national Workshop on Artificial Intelligence Safety Engineering (WAISE 2022), Sep 2022, München,
Germany. �hal-03769683�

HAL Id: hal-03769683

https://hal.science/hal-03769683
Submitted on 5 Sep 2022

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
Object Detection With Probabilistic Guarantees:
a Conformal Prediction Approach

Florence de Grancey1 , Jean-Luc Adam2 , Lucian Alecu3 , Sébastien

Gerchinovitz4,5 , Franck Mamalet5 , and David Vigouroux5
1
Thales AVS France SAS, Toulouse, France
2
Renault, Toulouse, France
3
Continental, Toulouse, France
4
Institut de Mathématiques de Toulouse, Toulouse, France
5
IRT Saint Exupéry, Toulouse, France
[email protected]

Abstract. Providing reliable uncertainty quantification for complex vi-

sual tasks such as object detection is of utmost importance for safety-
critical applications such as autonomous driving, tumor detection, etc.
Conformal prediction methods offer simple yet practical means to build
uncertainty estimations that come with probabilistic guarantees. In this
paper we apply such methods to the task of object localization and il-
lustrate our analysis on a pedestrian detection use-case. Throughout the
paper we highlight both theoretical and practical implications of our
analysis.

Keywords: Object detection · Conformal prediction · Uncertainty quan-

tification.

1 Introduction

Recent works in object detection show a great variety of models and approaches.
Among the most notable we can mention: RCNN [13], Fast-RCNN [12], Reti-
naNet [22], FPN [21], YOLO and its several versions [28–30], SSD [24] or DETR
[7].
Despite their impressive success observed on various benchmarks, many chal-
lenges remain ahead. For critical systems, several additional guarantees shall be
provided to avoid catastrophic consequences: in an autonomous vehicle, a pedes-
trian mislocated by the system could be hurt or killed; in a cancer detection
system, several cancer cells missed by the object detector could not be treated.
To ensure the safety of the user, the uncertainty of the location of the object to
detect should be quantified, allowing to create safeguards around the object.
The main challenge consists in providing reliable uncertainty quantification of
their prediction errors. While many object detection models compute so-called
confidence scores which can be interpreted as basic estimators of uncertainty,
they are often unreliable (i.e. over or under-estimating the true uncertainty).
2 F. de Grancey et al.

Another difficulty stems from the complex interplay between the classification-
type errors and the localization-type errors of the object detectors. In addition,
the risks associated with each type of error are application-dependent.
For safety-related applications, one may seek to obtain various guarantees.
One such guarantee related to object localization, that will be addressed in this
paper, may read: ensure that at least a significant portion (i.e. a user-specified
fraction) of the objects recognized in visual images satisfy this property: their
true bounding boxes are fully covered6 by the boxes predicted by a given object
detection model. This type of guarantee may be helpful, for example, to build
reliable models for tumor discovery, obstacle detection or trajectory estimation.

Uncertainty Quantification for Object Detection Several techniques such

as Deep Ensembles methods [18, 25] or Monte Carlo-Dropout methods [2, 9, 26,
27] have been developed to provide epistemic uncertainty quantification. Other
methods such as Direct Modeling add additional layers on top of the object
detector to achieve such estimations [19]. More recent works introduce proba-
bilistic object detectors which distinguish between the aleatoric and epistemic
uncertainties and estimate their variances separately. Based on the work of [15],
the authors extend Bayesian neural networks to object detectors [14,16]. A com-
plete survey on uncertainty estimation for object detectors can be found in [11].
Many of the cited methods have been applied to uncertainty quantification
in object detection tasks with various success. Nevertheless, to the best of our
knowledge, none of these works provide statistical guarantees about the esti-
mated uncertainties, e.g., that the relevant objects are correctly classified with
high probability, or that they are correctly localized in the image most of the
time, or both.

Main Contributions and Outline of the Paper In this work, we consider a

relatively recent family of statistical methods called Conformal Prediction, which
are post-processing methods to compute guaranteed “error margins” for various
learning tasks. Our main contribution is the first application of such ideas to a
practical object detection use-case, namely, pedestrian localization (i.e. correct
prediction of the minimum area bounding box encompassing objects classified as
pedestrians). This may be further used to increase the reliability of, e.g., collision
avoidance or assisted braking functions. To that end, the paper is organized as
follows:

– After presenting the main ingredients of conformal prediction in Section 2

we describe the experimental setting in Section 3.
– In Sections 4 and 5 we show several ways to apply conformal prediction
methods for object localization, with various statistical guarantees.
– In Section 6 we emphasize subtle pitfalls that a user may fall into, to help
interpret conformal prediction guarantees when applied to object detection.

6
All the coordinates of the true box will be found inside the rectangle defined by the
predicted bounding box of the object.
 Object Detection With Probabilistic Guarantees 3

Fig. 1. Conformalization example (box-wise, risk level α = 0.1) on a BDD100k image

with Ground Truth, Inference and Conformalized boxes.

2 Background: Conformal Prediction

Consider a supervised learning task (e.g. classification or regression), where we

want to predict an unknown label y (e.g. a class or a real number) given an ob-
served input x (e.g. an image). Typical ML models such as deep neural networks
output predictions fˆ(x) with little or no hint as to whether fˆ(x) is close to the
unknown label y. To that end, Conformal Prediction [1, 20, 33] is a family of
post-processing methods that are useful to compute guaranteed “error margins”,
under some assumptions on the data (see Theorem 1 below for an example). The
overall process from learning to inference typically unfolds as follows.7

1. Data collection: Two different datasets are collected: a training set and a
calibration set, which will be used to learn and evaluate a ML model. (See
below for independence and distribution requirements on the data.)
2. Training step: a machine learning model fˆ is learned on the training set.
The underlying model can be of virtually any kind (a deep neural network,
a random forest, etc).
3. Conformalization step: the learned model fˆ is evaluated on the calibration
set. This step consists in measuring the errors of fˆ on the calibration set,
and in reporting a quantile qα of these errors for some pre-specified

risk level
α ∈ (0, 1). More precisely, given a non-conformity score s ŷ, y to assess the
“distance” between a prediction ŷ = fˆ(x) and a ground truth y, we compute
the errors of fˆ on all data points (xi , yi ) of the calibration set8 :

Ri = s ŷi , yi ,


i = 1, . . . , nc , (1)
7
More complex variants exist. The typical process outlined here is more precisely
known as split conformal prediction.
8
The errors are sometimes called “residuals” (hence the Ri notation).
4 F. de Grancey et al.

where nc is the size of the calibration
set. (For example, in regression, we

can take the absolute difference s ŷ, y = |y − ŷ|.) Then, the quantile qα is
defined as the d(1−α)(nc +1)e-th largest value among the observed errors Ri .
4. Inference step: Given a new input x, instead of outputting a simple pre-
diction fˆ(x), we output a prediction set C α (x), with the goal of containing
the unknown ground truth y. It is defined as the set of all labels y 0 that are
“close enough” to the prediction ŷ = fˆ(x) of the ML model:
n o
C α (x) = y 0 : s ŷ, y 0 ≤ qα ,


(2)

where qα is the error quantile reported at the end of the conformalization

step, and serves as an “error margin”. For example, in the regression example
mentioned above, the prediction set is given by C α (x) = [ŷ−qα ; ŷ+qα ]. Other
non-conformity scores lead to other prediction sets, as shown later.

Dataset Requirements. In order to be able to prove that a prediction set

C α (x) contains the unknown label y “most of the time”, the datasets must satisfy
some requirements. Sufficient requirements are that:9
(i) data from all 3 datasets (training, calibration, inference) are independent;
(ii) data distributions at calibration and inference steps are identical.
Requirement (i) is useful to avoid overfitting issues. Requirement (ii) is useful
to make sure that errors measured during the conformalization step are rep-
resentative of errors at inference time. Interestingly though, training data can
be distributed differently, which can be useful when computational resources or
data for training are rather scarce, while an ML model carefully trained for a
close distribution is already available. (Of course, a model that was pre-trained
for a very different distribution will perform poorly at the conformalization step,
and thus the error margin qα will be large.)
Under the above dataset requirements, the conformal prediction process 1-4
outlined above satisfies the following probabilistic guarantee.
Theorem 1 (see, e.g., [1, 20, 33]). Assume the training, calibration, and in-
ference datasets satisfy Requirements (i) and (ii) above. Then, on average over
the choice of the calibration set and the new data point (X, Y ),

P Y ∈ C α (X) ≥ 1 − α .

We say in this case that the method has a coverage of 1 − α. The above
guarantee means that, for a fraction 1−α of all possible calibration sets in Step 3
and possible data points (x, y) in Step 4, the prediction set C α (x) contains the
true label y. In other words, if we repeated the overall conformal prediction
process 1-4 many times independently, it would err a fraction at most α of the
time. Details about dangers of interpretation are given in Section 6.
9
Mathematically speaking, it is in fact sufficient that the calibration data and the
data at inference time are exchangeable, conditionally on the training data.
 Object Detection With Probabilistic Guarantees 5

3 Experimental Setting and Goals

In the following sections we describe how conformal prediction methods can be
applied to post-process (i.e. shrink or enlarge) the prediction boxes provided
by a pedestrian detector. The goal is that the new boxes, called conformalized
boxes, fully cover the true bounding boxes of the objects of interest, “most of
the time”. For example, in Figure 1 we would like to cover all gold boxes with
green boxes, which are obtained by adjusting the predicted boxes in cyan. The
precise interpretation and limitations of the “most of the time” statement will be
detailed shortly, when we apply conformalization procedures on different levels
respectively: per coordinate, per bounding box or per image.
Both the level at which conformalization is conducted, and the design of non-
conformity scores are engineering choices, as they depend on the actual usage of
the model in real-world applications. For the pedestrian detection case considered
here we may be interested in providing guarantees related to individual objects.
In this case, a box-wise conformalization seems more appropriate. However, in
other cases we may be interested in image-wise conformalization, as to ensure
that a majority of images satisfy a desired property (e.g. all or most of the
objects of interest in the image are “well” localized).
In all our experimental settings we consider the YOLOv3 object detector [30]
pretrained on the COCO training dataset [23] (i.e. Step 2 described in Section 2
is fixed). As stated in Section 2 we can conformalize on a calibration dataset with
a distribution that is different from that of the training set. Therefore we conduct
all our experiments on the BDD100k dataset [34] by considering its training set
calib
as our calibration set (denoted by DBDD thereafter) and its validation set as our
test
test set (denoted by DBDD ). Since we focus on pedestrian detection, the original
70k +10k images of training and validation sets reduce to 22213 and 3220 images
with at least one person/pedestrian (we include riders), containing 91349 and
13262 annotated persons respectively.
In the following sections we propose non-conformity scores for each level of
analysis (coordinate, box or image) and discuss practical implications of these
choices. Finally, we emphasize some statistical aspects that are essential to any
correct interpretation of the obtained guarantees.

4 Coordinate-Wise and Box-Wise Conformalization

As mentioned above, our goal is to post-process the boxes predicted by an object
detector such that they cover the true bounding boxes of the objects of interest.
In this section, we compute error margins at box level, by treating boxes as indi-
vidual data points for the conformal prediction process of Section 2. We compute
error margins for each coordinate xmin , xmax , ymin , ymax separately (Sec. 4.2), and
then show how to correct them to obtain guarantees at box level (Sec. 4.3).

4.1 Preliminary Assignment

On a given image, in order to compare predicted boxes with true boxes at box
level (that is, compare box A with box B), we need to assign predicted boxes to
6 F. de Grancey et al.

true boxes. A preliminary pre-processing assignment stage is thus necessary. This

step is performed with the Hungarian matching algorithm [17] based on the IoU
metric. By applying it, we retain exclusively the true positive bounding boxes
for calibration. Therefore, at inference time our probabilistic guarantee will only
ensure that true positives are correctly covering the ground truth, while false
negatives might still exist. In the following experiments, unless otherwise stated,
a detection box must have a confidence score higher than 0.5 and an IoU with a
ground truth object above 0.5 to assign the predicted box to a true box. Based
on this assignment stage, the BDD100k calibration and test sets reduce to 42824
and 6138 assigned persons respectively.

4.2 Coordinate-Wise Conformalization

We explain in details how to instantiate Steps 3 and 4 of Section 2.

Conformalization step. Assume that we have assigned predicted boxes to true

boxes as in Section 4.1. In order to compare the i-th predicted box with the i-th
true box, we compare each of the four predicted coordinates x̂imin , x̂imax , ŷmin
i i
, ŷmax
i i i i
with the four true coordinates xmin , xmax , ymin , ymax , by counting errors posi-
tively when the truth lies outside the prediction (e.g. x̂imin > ximin or x̂imax <
ximax ), and negatively otherwise. This choice is less conservative than consider-
ing absolute error values and leads to the following four errors (cf. Eq. (1)): for
i = 1, . . . , nc ,

Rxi min = x̂imin − ximin i

Ryi min = ŷmin i
− ymin
(3)
Rxi max = ximax − x̂imax i i i
Rymax = ymax − ŷmax

Note that nc is given by the total number of predicted objects assigned to a

true object, which is larger than the number of images in the calibration set.
Then, following Step 3 of Section 2, we compute a quantile qα for each of
the four errors above, defined as the d(1 − α)(nc + 1)e-th largest value among
the observed errors Ri . These four quantiles will serve as error margins for each
coordinate.
As an illustration, the errors Ryi max on DBDD
calib
are represented on the histogram
on the left side of Fig. 2 (in red: the quantile qα for α = 0.1, i.e. specified coverage
of 0.9). The right side shows the evolution of the quantile qα w.r.t. the parameter
α. High guarantees (α < 0.05) imply large margins, whereas low guarantees only
require small modifications of the predicted coordinates.

Inference step. We now instantiate Eq. (2) of Section 2 to compute a pre-

diction set C α for each coordinate xmin , xmax , ymin , ymax , given four predicted
coordinates x̂min , x̂max , ŷmin , ŷmax as inputs. These prediction sets are intervals:

Cxαmin = [x̂min − qαxmin , +∞) Cyαmin = [ŷmin − qαymin , +∞)

(4)
Cxαmax = (−∞, x̂max + qαxmax ] Cyαmax = (−∞, ŷmax + qαymax ]
 Object Detection With Probabilistic Guarantees 7

quantile alpha=0.1
mean 25

4000
20

Quantile value in Pixels

3000
15

2000 10

5
1000

0
0
−20 −15 −10 −5 0 5 10 15 20 0.0 0.1 0.2 0.3 0.4 0.5 0.6
Errors Ri in pixels Risk Level alpha

Fig. 2. Left: histogram of the errors Ri for the coordinate ymax , and the corresponding
quantile qα for α = 0.1. Right: Evolution of the quantile value qα with risk level α.

In Table 1, the first four lines (coordinate-wise) give the evaluation of the
test
observed coverage on the DBDD set, i.e., for each coordinate, the proportion of
(true positive) boxes for which the true coordinate lies within the corresponding
prediction set. We can see that Thm. 1 is verified whatever the specified coverage.

Table 1. Evaluation of observed coverage on DBDD test

using the quantile evaluated for
three specified coverage values (in red when the specified coverage is not reached).

Method specified coverage (1-α) 0.7 0.9 0.95

Observed coverage
xmin 0.76 0.91 0.96
Coordinate-Wise xmax 0.78 0.91 0.96
§4.2 ymin 0.70 0.92 0.95
ymax 0.71 0.91 0.95
Box-Wise §4.3 w/o Bonferroni 0.35 0.73 0.86
Box-Wise §4.3 with Bonferroni 0.79 0.92 0.96

4.3 Bonferroni Correction for Box-Level Guarantees

In this section we seek the following guarantee: at inference time, among all
true bounding (pedestrian) boxes that are detected, a fraction 1 − α of them
are correctly covered by conformalized boxes.10 We explained in Eq. (4) how to
compute error margins to locate unknown coordinates xmin , xmax , ymin , ymax of a
box, given predictions x̂min , x̂max , ŷmin , ŷmax . It might be tempting to define the
10
The 1 − α guarantee only holds on average over all calibration sets, see Section 6.
8 F. de Grancey et al.

conformalized box as the largest (worst-case) box whose coordinates are within
the intervals C α , i.e., the box with coordinates

x̂min − qαxmin x̂max + qαxmax ŷmin − qαymin ŷmax + qαymax .

However, looking at the experimental results of Table 1 (line w/o Bonferroni),

we remark that for, e.g., a specified coverage of 95%, only 86% of true boxes are
covered by conformalized boxes. The lower-than-expected coverage is a direct
consequence of considering each coordinate of the bounding boxes independently,
since errors on each coordinate can happen on different boxes.
One simple solution is to apply a so-called Bonferroni correction (e.g., [5]), a
statistical adjustment to account for this 4 − way dependency. This is straight-
forward, as it amounts to perform the exact same procedure as described in the
previous section, but replacing qα with qα/4 . With this slight change the new
coverage on the test set becomes within the expected levels (see Table 1, line
with Bonferroni). Note that since qα/4 ≥ qα , the conformalized boxes will be
larger. For example, for the ymax coordinate and for a specified coverage of 0.9
(α = 0.1), we obtain qα = 6 pixels and qα/4 = 14 pixels.

5 Image-Wise Conformalization

The method of the previous section aimed at guaranteeing that, at inference

time, among all true (pedestrian) bounding boxes that are detected, a fraction
1 − α of them are correctly covered by conformalized boxes.10 This guarantee
has two limitations: (i) false negatives (undetected pedestrians) are not taken
into account, without any control on their occurrence rate; (ii) the fraction of
detected true boxes that are covered on a given image might be much different
from 1 − α (the 1 − α coverage is an average over all boxes across the test set).
While the box-wise approach can still be useful for pedestrian detection (e.g., for
tracking true positives), for medical applications such as cancer cell detection,
(i) and (ii) imply that we might miss too many cancer cells on too many images.
Next we study another non-conformity score to pursue a guarantee at image
level rather than at object level. We aim at the following guarantee: on average
over the choice of the calibration set, a fraction 1 − α of images at inference will
be such that a fraction 1 − β of true boxes in the image will be correctly covered
by conformalized boxes.
To that end, we consider a non-conformity score sβ (B̂, B) that is a close
variant of the partial (or quantile) directed Hausdorff distance [31]. It compares
two sets of boxes on a given image: the set B̂ of all predicted boxes, and the set
B of all true boxes. Our score sβ (B̂, B) is defined as the smallest margin r ≥ 0
that it suffices to add to all predicted boxes in B̂ (on all four coordinates) so
that a fraction at least 1−β of true boxes in B are correctly covered by the union
of enlarged predicted boxes. We then follow the whole process of Section 2. In
particular, the quantile qα computed at conformalization is the margin that will
be added to all predicted boxes at inference.
 Object Detection With Probabilistic Guarantees 9

Fig. 3. Conformalization example (image-wise, risk level α = 0.1, β = 0.25) on a

BDD100k image with Ground Truth, Inference and Conformalized boxes.

A BDD100k example is shown on Figure 3, for α = 0.1 and β = 0.25 (the

aim is to correctly locate 75% of all pedestrians on each of 90% of all possi-
ble images). Interestingly, though only 5 pedestrians were detected (in cyan),
9 pedestrians (in gold) are located within conformalized boxes (in green). This
positive effect is due to the presence of multiple pedestrians nearby. Though this
effect is not frequent in our use-case (pedestrians are often more isolated) and
prevents the system from tracking pedestrians individually, it seems more useful
for applications where objects of interest are often nearby, such as cancer cell
detection.

6 Statistical Pitfalls

While very useful in practice, the probabilistic guarantees behind conformal

prediction such as Theorem 1 should be interpreted with care. They rely on
assumptions and have some limitations, which we outline below.
A guarantee “on average” over the box/image domain. As explained
after Theorem 1, the inequality P (Y ∈ C α (X)) ≥ 1 − α is a guarantee on
average over calibration sets and test data points. In our use-case, this implies
that the 1 − α coverage is correct on average over all boxes/images at inference
time, but might be incorrect in some subsets of the box/image domain. For
example, for our box-wise conformalization experiment on the BDD100k dataset,
test
when restricting the test set DBDD to pedestrians that are close to the camera
(height larger than 150 pixels), the coverage is smaller than 64% for a specified
coverage of 70% (α = 0.3), and smaller than 81% for a specified coverage of 90%
(α = 0.1).11 Of course, a simple solution here is to apply conformal prediction
for close pedestrians only (predicted height larger than 150 pixels). However, this
11
These coverage values include statistical error margins at level 95%.
10 F. de Grancey et al.

solution cannot work in settings where a very large number of subdomains need
to be distinguished (since calibration sets need to be large enough) or where
these subdomains are not known a priori. In any case, due to the statistical
nature of conformal prediction methods, one must keep in mind that there are
some boxes or images on which these methods will fail at inference.
A guarantee “on average” over calibration sets. Similarly, while the
1 − α coverage is correct on average over all possible calibration sets, its value
might be different for the single calibration set used in practice. The way the
coverage varies from one calibration set to another was described in details in [1].
Next we illustrate this variability on BDD100k with box-wise conformalization
(as in Section 4.3) and α = 0.1, by re-sampling various calibration sets and re-
porting the associated test coverage values. The histogram on Figure 4 shows a
large variability of coverage values, which means that different calibration sets
lead to different coverage values at inference. Fortunately here, most values are
above the specified coverage of 0.9 (since the margins are a little conservative
due to the Bonferroni correction), but the tail probability on the left of 0.9 im-
plies that the user has still a chance to use a calibration set that would result in
a lower-than-expected coverage at inference. Recent works have proposed vari-
ants of conformal prediction (with more conservative margins) to deal with this
variability (e.g., [4, 10]).

Empirical coverage at box level, for a risk level 0.1

40

35

30

25

20

15

10

0
0.88 0.90 0.92 0.94 0.96
Fraction of Valid Boxes

Fig. 4. Empirical coverage distribution measured on the same test set when sampling
different calibration sets and applying box-wise conformalization as in Section 4.3.

Datasets requirements: independence assumption. As recalled in Sec-

tion 2, several properties on the datasets are required for the probabilistic guar-
antee of Theorem 1 to be valid. This is not at all surprising due to the statistical
 Object Detection With Probabilistic Guarantees 11

nature of conformal prediction. The independence requirement (i) between all

data involved (training, calibration, test) is to avoid overfitting issues. In partic-
ular, dependencies between training and calibration data could lead to measure
artificially small errors Ri at conformalization, while dependencies within the
calibration set may reduce statistical power in an uncontrolled manner.12 In
practice, the independence assumption seems difficult to guarantee or test, but
data collection practices should aim at it. For the box-wise procedure of Sec-
tion 4.3, it is not clear a priori that errors Ri are independent, because of pos-
sible intra-image dependencies between boxes. Dependencies seem a little easier
to prevent in the image-wise setting of Section 5 (by, e.g., discarding neighboring
images in a sequence). For our use-case however, our box-wise and image-wise
test coverage values, which reach specified coverage, are really encouraging.
Distribution shift between calibration and test. We now investigate
the importance of the dataset requirement (ii) of Section 2. To that end, we use
calib
our model calibrated on DBDD for the box-wise criterion, as in Section 4.3, but we
test its coverage on a new dataset, namely the Cityscape validation set [8]. This
might be a reasonable thing to do, as both datasets are quite similar (they include
urban scenes for autonomous driving scenarios). Yet the obtained coverage on
this new dataset is 0.790 for α = 0.1, i.e. significantly lower than expected.
test
On the opposite, the coverage on DBDD is 0.924, which is consistent with the
expected guarantee. See Figure 5 for results on a range of α values, the black
dotted line showing the expected coverage. In our haste to apply our methods on
new data, we misinterpreted the offered guarantee: even if conceptually similar,
the two datasets (the one used for calibration and test) probably do not share the
same intrinsic data distribution. This invalidates the data-related assumptions
described in Section 2, and consequently the offered guarantee. While this is
obviously a contrived example, in many practical situations it is quite easy to
mislead oneself into believing that the data processed at inference time follows
exactly the same distribution as the one used for the calibration of the model.
Recent works in conformal methods such as [3, 32] attempt to address such
challenges related to distribution shift. In our future work, we aim to apply these
results on our own experiments.

7 Discussions and Future Works

In this paper we present a practical guideline of how conformal prediction meth-
ods can be applied to object localization to provide uncertainty estimations with
probabilistic guarantees. We illustrate the essential aspects on a pedestrian lo-
calization use case. We propose several variants of conformal prediction methods
that provide guarantees at various levels (coordinate, bounding box or image)
and with different non-conformity scores. Finally, we highlight several statistical
aspects that one must take into account, and discuss the interpretation of the
obtained guarantees.
12
In fact, conformal guarantees work slightly beyond independence—under a so-called
“exchangeability” assumption [1, 20, 33].
12 F. de Grancey et al.

1.0 BDD100k Validation Set 1.0 Cityscape Validation Set 120 Conformalized with BDD100k
Conformalized with Cityscape
0.9 0.9 100

0.8 0.8 80

Pixels
0.7 0.7 60

0.6 0.6 40

0.5 0.5 20

0.4 0.4 0
0.0 0.2 0.4 0.6 0.0 0.2 0.4 0.6 0.0 0.2 0.4 0.6
Risk Level alpha Risk Level alpha Risk Level alpha

Fig. 5. Box-wise empirical coverage oberved on (left) BDD100k validation set vs.
(middle) Cityscape validation set (expected coverage is black dotted line). Right: quan-
tile curves for Ymax with BDD100k or Cityscape training sets considered for calibration.

An engineering choice. Conformal prediction methods rely on the choice of

a non-conformity score that can be designed according to the targetted task and
a priori knowledge. For example, if we wish to differentiate the level of uncertain-
ties w.r.t. the size of the object we can normalize the scores of Eq. 3 by the width
and height of the predicted box, i.e., R̃xi min = Rxi min /∆ix , R̃xi max = Rxi max /∆ix ,
where ∆ix = |x̂imax − x̂imin | (and similarly for the y coordinates). Such scores
would lead to margins that scale multiplicatively with the size of the predicted
object. Though this is undesirable for our pedestrian use-case (since pedestri-
ans that are close to the camera correspond to the largest objects), this might
be useful in applications where we are ready to pay larger margins for larger
objects. Likewise, we can extend proposed scores as to compute simultaneously
an inner and outer conformal bounding box covering most of the true boxes (by
considering errors in absolute value). Note that the choice of the non-conformity
score can also help to analyze the data quality in test or calibration sets: high
scores can be indicative of ’extreme cases’, such as suspicious inputs or anoma-
lous annotations.
Future work. This paper focused on the localization task. Next we will
address the classification problem, and the interplay between the two—that is,
the global detection problem. At a system level, it would also be important to
build a link between such statistical guarantees and safety-related risks (e.g.,
fault rate). This task is however difficult, as this necessarily relies on additional
assumptions and uncertainties [6]. All these aspects, together with the numerous
potential statistical pitfalls, open up very challenging research questions.

Acknowledgements This work has benefited from the AI Interdisciplinary

Institute ANITI, which is funded by the French “Investing for the Future –
PIA3” program under the Grant agreement ANR-19-P3IA-0004. The authors
gratefully acknowledge the support of the DEEL project.13
13
https://www.deel.ai/
 Object Detection With Probabilistic Guarantees 13

References
1. Angelopoulos, A.N., Bates, S.: A gentle introduction to conformal prediction and
distribution-free uncertainty quantification (2021), arXiv:2107.07511
2. Azevedo, T., de Jong, R., Maji, P.: Stochastic-yolo: Efficient probabilistic object
detection under dataset shifts (2020), arXiv:2009.02967
3. Barber, R.F., Candes, E.J., Ramdas, A., Tibshirani, R.J.: Conformal prediction
beyond exchangeability (2022), arXiv:2202.13415
4. Bates, S., Angelopoulos, A., Lei, L., Malik, J., Jordan, M.I.: Distribution-free,
risk-controlling prediction sets. Journal of the ACM 68(6) (2021)
5. Bickel, P.J., Doksum, K.A.: Mathematical Statistics: Basic Ideas and Selected Top-
ics, vol. 1. Chapman and Hall/CRC (2015)
6. Bonnin, H., Jenn, E., Alecu, L., Fel, T., Gardes, L., Gerchinovitz, S., Ponsolle,
L., Mamalet, F., Mussot, V., Cappi, C., Delmas, K., Lefevre, B.: Can we reconcile
safety objectives with machine learning performances? In: Proc. of the 11th Edition
of European Congress of Embedded Real Time Systems (ERTS) (2022)
7. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-
to-end object detection with transformers. In: Computer Vision – ECCV 2020.
Springer International Publishing (2020)
8. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R.,
Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene
understanding. In: 2016 IEEE Conference on Computer Vision and Pattern Recog-
nition (CVPR) (2016)
9. Deepshikha, K., Yelleni, S.H., Srijith, P.K., Mohan, C.K.: Monte carlo dropblock
for modelling uncertainty in object detection (2021), arXiv:2108.03614
10. Ducoffe, M., Gerchinovitz, S., Sen Gupta, J.: A high-probability safety guarantee
for shifted neural network surrogates. In: Proceedings of the Workshop on Artificial
Intelligence Safety (SafeAI 2020). pp. 74–82 (2020)
11. Feng, D., Harakeh, A., Waslander, S.L., Dietmayer, K.: A review and comparative
study on probabilistic object detection in autonomous driving. IEEE Transactions
on Intelligent Transportation Systems (2021)
12. Girshick, R.B.: Fast r-cnn. In: 2015 IEEE International Conference on Computer
Vision (ICCV) (2015)
13. Girshick, R.B., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for
accurate object detection and semantic segmentation. In: 2014 IEEE Conference
on Computer Vision and Pattern Recognition (2014)
14. Harakeh, A., Smart, M., Waslander, S.L.: BayesOD: A Bayesian approach for un-
certainty estimation in deep object detectors (2019), arXiv:1903.03838
15. Kendall, A., Gal, Y.: What uncertainties do we need in bayesian deep learning for
computer vision? In: Proceedings of the 31st International Conference on Neural
Information Processing Systems (2017)
16. Kraus, F., Dietmayer, K.: Uncertainty estimation in one-stage object detection. In:
2019 IEEE Intelligent Transportation Systems Conference (ITSC) (2019)
17. Kuhn, H.W.: The hungarian method for the assignment problem. Naval Research
Logistics Quarterly 2(1-2) (1955)
18. Lakshminarayanan, B., Pritzel, A., Blundell, C.: Simple and scalable predictive
uncertainty estimation using deep ensembles. In: Proceedings of the 31st Interna-
tional Conference on Neural Information Processing Systems (2017)
19. Le, M.T., Diehl, F., Brunner, T., Knol, A.: Uncertainty estimation for deep neural
object detectors in safety-critical applications. In: 2018 21st International Confer-
ence on Intelligent Transportation Systems (ITSC) (2018)
14 F. de Grancey et al.

20. Lei, J., G’Sell, M., Rinaldo, A., Tibshirani, R.J., Wasserman, L.: Distribution-free
predictive inference for regression. Journal of the American Statistical Association
113(523), 1094–1111 (2018)
21. Lin, T., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyra-
mid networks for object detection. In: 2017 IEEE Conference on Computer Vision
and Pattern Recognition (CVPR) (2017)
22. Lin, T., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object
detection. IEEE Transactions on Pattern Analysis Machine Intelligence 42(02)
(2020)
23. Lin, T., Maire, M., Belongie, S.J., Bourdev, L.D., Girshick, R.B., Hays, J., Perona,
P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft COCO: common objects in
context. In: Computer Vision – ECCV 2014. Springer International Publishing
(2014)
24. Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S.E., Fu, C., Berg, A.C.: Ssd:
Single shot multibox detector (2016)
25. Lyu, Z., Gutierrez, N., Rajguru, A., Beksi, W.J.: Probabilistic object detection via
deep ensembles. In: European Conference on Computer Vision. Springer (2020)
26. Miller, D., Dayoub, F., Milford, M., Sunderhauf, N.: Evaluating merging strategies
for sampling-based uncertainty techniques in object detection. In: 2019 Interna-
tional Conference on Robotics and Automation (ICRA) (2019)
27. Miller, D., Nicholson, L., Dayoub, F., Sünderhauf, N.: Dropout sampling for ro-
bust object detection in open-set conditions. In: 2018 International Conference on
Robotics and Automation (ICRA) (2018)
28. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified,
real-time object detection. In: 2016 IEEE Conference on Computer Vision and
Pattern Recognition (CVPR) (2016)
29. Redmon, J., Farhadi, A.: Yolo9000: Better, faster, stronger. In: 2017 IEEE Con-
ference on Computer Vision and Pattern Recognition (CVPR) (2017)
30. Redmon, J., Farhadi, A.: Yolov3: An incremental improvement (2018),
arXiv:1804.02767
31. Rucklidge, W.J.: Efficiently locating objects using the hausdorff distance. Interna-
tional Journal of Computer Vision 24, 251–270 (1997)
32. Tibshirani, R.J., Barber, R.F., Candes, E.J., Ramdas, A.: Conformal prediction un-
der covariate shift. In: Proceedings of the 33rd International Conference on Neural
Information Processing Systems (2019)
33. Vovk, V., Gammerman, A., Shafer, G.: Algorithmic Learning in a Random World.
Springer-Verlag, Berlin, Heidelberg (2005)
34. Yu, F., Xian, W., Chen, Y., Liu, F., Liao, M., Madhavan, V., Darrell, T.:
BDD100K: a diverse driving dataset for heterogeneous multitask learning (2018),
arXiv:1805.04687