PR-207: YOLOv3: An Incremental Improvement

TensorFlow Korea 논문읽기모임 PR12 207번째 논문 review입니다

이번 논문은 YOLO v3입니다.
매우 유명한 논문이라서 크게 부연설명이 필요없을 것 같은데요, Object Detection algorithm들 중에 YOLO는 굉장히 특색있는 one-stage algorithm입니다. 이 논문에서는 YOLO v2(YOLO9000) 이후에 성능 향상을 위하여 어떤 것들을 적용하였는지 하나씩 설명해주고 있습니다. 또한 MS COCO의 metric인 average mAP에 대해서 비판하면서 mAP를 평가하는 방법에 대해서도 얘기를 하고 있는데요, 자세한 내용은 영상을 참고해주세요~

논문링크: https://arxiv.org/abs/1804.02767
영상링크: https://youtu.be/HMgcvgRrDcA

1. YOLOv3:
An Incremental Improvement
Joseph Redmon, et al., “YOLOv3: An Incremental Improvement”
17
th
November, 2019
PR12 Paper Review
JinWonLee
Samsung Electronics

2. Related PR12
•PR-016: You only look once: Unified, real-time object detection :
https://youtu.be/eTDcoeqj1_w
•PR-023: YOLO9000: Better, Faster, Stronger :
https://youtu.be/6fdclSGgeio

3. References
•What’s new in YOLO v3?
https://towardsdatascience.com/yolo-v3-object-detection-
53fb7d3bfe6b
•How to implement a YOLO (v3) object detector from scratch in
PyTorch
https://blog.paperspace.com/how-to-implement-a-yolo-object-
detector-in-pytorch/

4. IOU & mAP
mAP

5. Introduction
•This is a TECH REPORT.
•“I managed to make some improvements to YOLO. But, honestly,
nothing like super interesting, just a bunch of small changes that
make it better.”
•Better, Not Faster, Stronger(?)
•The last sentence of the document: In closing, do not @me.
(Because I finally quit Twitter).

6. Bounding Box Prediction
•YOLOv3 predicts bounding boxes using dimension clusters as anchor
boxes. The network predicts 4 coordinates for each bounding box, t
x,
t
y, t
w, t
h.
•Use sum of squared error loss

7. Bounding Box Prediction
•YOLOv3 predicts an objectnessscore for each bounding box using
logistic regression.
•This should be 1 if the bounding box prior overlaps a ground truth
object by more than any other bounding box prior.
•They use the threshold of 0.5. Unlike Faster R-CNN, YOLOv3 only
assigns one bounding box prior for each ground truth object.

8. Class Prediction
•YOLO v3 now performs multilabelclassification for objects detected
in images.
•Softmaxingclasses rests on the assumption that classes are mutually
exclusive.
•However, when we have classes likePersonandWomenin a dataset,
then the above assumption fails. This is the reason why the authors
of YOLO have refrained from softmaxingthe classes. Instead, each
class score is predicted using logistic regression and a threshold is
used to predict multiple labels for an object.

9. Predictions Across Scales
•The most salient feature of YOLOv3 is that
it makes detections at three different scales.
•YOLOv3 predicts boxes at 3 scales
•YOLOv3 predicts 3 boxes at each scale
in total 9 boxes
So the tensor is N x N x (3 x (4 + 1 + 80))
80
3
N
N
255

10. Anchor Boxes
•They still use k-means clustering to determine bounding box priors.
They just sort of chose 9 clusters and 3 scales arbitrarily and then
divide up the clusters evenly across scales.
•On the COCO dataset the 9 clusters were:
(10x13), (16x30), (33x23), (30x61), (62x45), (59x119), (116x90), (156x
198), (373x326).

11. No. of Bounding Boxes
•YOLOv1predicts 98 boxes (7x7 grid cells, 2 boxes per cell @448x448)
•YOLOv2 predicts 845 boxes (13x13 grid cells, 5 anchor boxes
@416x416)
•YOLOv3 predicts 10,647 boxes (@416x416)
•YOLOv3 predicts more than 10x the number of boxes predicted by
YOLOv2

12. Feature Extraction
•Darknet-53
•Darknet-53 is better than ResNet-101 and 1.5x
faster. Darknet-53 has similar performance to
ResNet-152 and is 2x faster.
•Darknet-53 also achieves the highest
measured floating point operations per
second. This means the network structure
better utilizes the GPU, making it more
efficient to evaluate and thus faster.

13. YOLOv3 Architecture
Darknet-53
Similar to Feature Pyramid Network

14. Training
•Authors still train on full images with no hard negative mining or any
of that stuff.
•They use multi-scale training, lots of data augmentation, batch
normalization, all the standard stuff.

15. Results
•It is still quite a bit behind other models like RetinaNetin this metric
though.
•However, when we look at the “old” detection metric of mAPat IOU=
0.5 (or AP50 in the chart) YOLOv3 is very strong.
•With the new multi-scale predictions YOLOv3 has relatively high APs
performances.

16. Results

17. Things We Tried That Didn’t Work
•Anchor box x, y offset predictions
Using the normal anchor box prediction mechanism where you predict the x,
y offset as a multiple of the box width or height
•Linear x, y predictions instead of logistic
•Focal loss
Focal loss dropped YOLOv3’s mAPabout 2 points
•Dual IOU thresholds and truth assignment
Faster R-CNN uses two IOU thresholds during training. If a prediction
overlaps the ground truth by 0.7 it is as a positive example, by [0.3-0.7] it is
ignored, less than 0.3 for all ground truth objects it is a negative example.

18. What This All Means
•YOLOv3 is a good detector. It’sfast, it’saccurate. It’s not as great on
the COCO average AP between .5 and .95 IOU metric. But it’svery
good on the old detection metric of .5 IOU.
•Russakovskyet al report that that humans have a hard time
distinguishing an IOU of .3 from .5!
Training humans to visually inspect a bounding box with IOU of 0.3 and
distinguish it from one with IOU 0.5 is surprisingly difficult.

19. Rebuttal
•Graphs have not one but two non-zero origins

20. Rebuttal
•For PASCAL VOC, the IOU threshold was “set deliberately low to
account for inaccuracies in bounding boxes in the ground truth data”
•COCO can have better labelling than VOC since COCO has
segmentation masks. But there was the lack of justification for
updating mAP.
•Emphasis must mean it de-emphasizes something else, in this case
classification accuracy. A miss-classified example is much more
obvious than a bounding box that is slightly shifted.

21. mAP’sProblems
•They are both perfect! (mAP= 1.0)

22. A New Proposal
•mAPis screwed up because all that matters is per-class rank ordering.
•What about getting rid of per-class AP and just doing a global
average precision?
•Or doing an AP calculation per-image and averaging over that?
•Boxes are stupid anyway though, I’m probably a true believer in
masks except I can’t get YOLO to learn them.

23. Thank you