Digging into Sample Assignment Methods for Object Detection Hiroto Honda Oct. 1, 2020 

Hiroto Honda - Mobility Technologies Co., Ltd. (Japan) - homepage: https://hirotomusiker.github.io/ - blogs: Digging Into Detectron 2 - kaggle master: 6th place at Open Images Challenge ‘19 - Interests: Object Detection, Human Pose Estimation, Image Restoration About Me 

How to define training samples of object detection for the given feature map and ground-truth boxes Today I talk about.. 

Accuracy and Inference Time Comparison among Object Detectors because It’s Hard to See the Difference between Sampling Methods in a Fair Way Today I don’t talk about... 

Object Detection Input: Image Output: Bounding Boxes (xywh + class id + confidence) from: [H1] 

Example of 2-stage Detector [H1][3] : Faster R-CNN [1] + Feature Pyramid Network [2] How Object Detection Works 

Object Detectors Decomposed backbone dense head roi head neck every grid cell is responsible for detection recognition of one object from one ROI feature map 

Object Detectors Decomposed backbone dense head roi head neck every grid cell is responsible for detection 2-stage detector 1-stage (single-shot) detector recognition of one object from one ROI feature map 

Object Detectors Decomposed detector name backbone neck dense head roi head Faster R-CNN [1] w/ FPN[2] ResNet FPN RPN Fast RCNN Mask R-CNN [4] ResNet FPN RPN Mask RCNN RetinaNet [5] ResNet FPN RetinaNetHead - EfficientDet [6] EfficientNet BiFPN RetinaNetHead - YOLO [7-11] darknet etc YOLO-FPN YOLO layer - SSD [12] VGG - SSDHead - 2-stage detector 1-stage (single-shot) detector 

How are Feature Maps and Ground Truth Associated? from: [H1] 

Region Proposal Network detector name backbone neck dense head roi head Faster R-CNN [1] w/ FPN[2] ResNet FPN RPN Fast RCNN Mask R-CNN [4] ResNet FPN RPN Mask RCNN 

Region Proposal Network (RPN) INPUT OUTPUT from: [H1] 

visualization of an objectness channel (corresponding to one of three anchors) Multi-Scale Detection Results (objectness) stride = 4 stride = 8 stride = 64 stride = 32 stride = 16 from: [H1] 

Anchors three anchors per scale aspect ratio : (1,1), (1, 2), (2, 1) from: [H1] 

Grid cells at the coarse scale have large anchors = responsible for detecting large objects Anchors on Each Grid Cell from: [H1] 

How are Feature Maps and Ground Truth Associated? Answer: Define the ‘foreground grid cells’ by matching ‘anchors’ with GT boxes from: [H1] 

IoU = A ∩ B / A ∪ B Intersection Over Union (IoU) A B IoU=0.15 IoU=0.95 

IoU Matrix for Anchor-GT Matching 0 0 0.61 0.28 0 0 0 0 0 0 0 0 0 0 0 0.98 0 0 GT box 0 GT box 1 matched with GT box 1, foreground background ignored foreground (IoU ≧ T1) -> objectness target=1, regression target background (IoU < T2) : objectness target=0, no regression loss ignored (T2 ≦ IoU < T1) from: [H1] T1 and T2: predefined threshold values IoU value anchors position 0 position 1 position 2 

Sample Assignment of RPN 

RPN learns relative size and location between GT boxes and anchors Box Regression After Sample Assignment from: [H1] Δx =(x-xa)/wa) Δy =(y-ya)/ha Δw = log(w/wa) Δh = log(h/ha) 

RetinaNet / EfficientDet detector name backbone neck dense head roi head RetinaNet [5] ResNet FPN RetinaNetHead - EfficientDet [6] EfficientNet BiFPN RetinaNetHead - 

RetinaNet stem C2 C3 C4 C5 Input Image BGR, H, W P6, P7 P5 P4 P3 P2 + + + + Backbone RetinaNetHead cls_subnet -> cls_score bbox_subnet -> bbox_pred 

EfficientDet stem C2 C3 C4 C5 Input Image BGR, H, W P6, P7 P5 P4 P3 P2 + + + + Backbone RetinaNetHead cls_subnet -> cls_score bbox_subnet -> bbox_pred Backbone: EfficientNet Neck: BiFPN 

same as RPN - number of anchors and IoU thresholds are different Sample Assignment of RetinaNet and EfficientDet foreground (IoU ≧ T1) : class target = one-hot, regression target background (IoU < T2): class target = zeros, no regression loss ignored (T2 ≦ IoU < T1) [only RetinaNet] T1 and T2: predefined threshold values architecture num. anchors at grid cell T1 T2 Faster R-CNN 3 0.7 0.3 RetinaNet 9 0.5 0.4 EfficientDet 3 0.5 0.5 0 0 0.41 0.28 0 0 0 0 0 0 0 0 0 0 0 0.68 0 0 GT box 0 GT box 1 matched with GT box 1, foreground background ignored IoU value anchors position 0 position 1 position 2 [3] 

YOLO v1 / v2 / v3 / v4 / v5 detector name backbone neck dense head roi head YOLO [7-11] darknet etc YOLO-FPN YOLO layer - 

What makes YOLO is the YOLO layer YOLO detector P5 P4 P3 YOLO Layer bbox, class score, confidence darknet53 YOLOv3 architecture 

Sample Assignment of YOLO v2 / v3 0 0 0.38 0.18 0 0 0 0 0 0 0 0 0 0 0 0.98 0 0 GT box 0 GT box 1 ・・・ matched with GT box 1, foreground background matched with GT box 0, foreground ignored (IoU between prediction and GT > T1) only one anchor is assigned to one GT max max foreground (max-IoU) : objectness = 1. regression target background (other than max-IoU anchors): objectness = 0, no regression loss for the details, see [H2] T1: predefined threshold values position 0 position 1 position 2 anchors 

Sample Assignment of YOLO v4 / v5 0 0 0.88 0.78 0 0 0 0 0 0 0 0 0 0 0 0.98 0 0 foreground (v4: IoU > T1, v5: box w, h ratio < Ta ) : objectness = 1. regression target background (v4: IoU > T1, v5: box w, h ratio > Ta) : objectness = 0, no regression loss ignored (IoU > T2, only YOLOv4) GT box 0 GT box 1 ・・・ matched with GT box 1, foreground matched with GT box 0, foreground multiple anchors can be assigned to one GT anchors position 0 position 1 position 2 

YOLOv5 assigns three feature points for one target center -> higher recall Sample Assignment Comparison - YOLOv3 vs YOLOv5 see my kaggle discussion topic for the YOLOv5 details https://www.kaggle.com/c/global-wheat-detection/discussion/172436 

target assignment is so different between YOLO versions - which one is the best? Sample Assignment of YOLO series version scale num. anchors per scale assignment method assigned anchors per GT YOLO v1 1 0 center position comparison single YOLO v2 1 9 IoU comparison single YOLO v3 3 3 IoU comparison single YOLO v4 3 3 IoU comparison multiple YOLO v5 3 3 box size comparison additional neighboring 2 cells multiple 

“Anchor-Free” Detectors detector name backbone neck dense head roi head FCOS [13] ResNet FPN FCOSHead - CenterNet (objects as points) [14] Hourglass CenterNetHead - 

- Assign all the grid cells that fall into the GT box - only at the appropriate scale - ‘Center-ness’ score is used additionally to suppress low-quality predictions far from the GT center FCOS 

- objectness (center) target: heatmap with Gaussian kernels around GT centers - regression target assignment: one grid cell + surrounding points (optional) Objects as Points (CenterNet) 

Adaptive Sample Selection detector name backbone neck dense head roi head ATSS [15] ResNet FPN ATSSHead - 

Adaptive Sample Selection Adaptively define IoU threshold for each GT box IoUthreshold = mean(IoUs) + std(IoUs) sample candidates : K=9 nearby anchors from the GT center Improves performance of both anchor-based and anchor-free detectors 

Adaptive Sample Selection 0 0 0.88 0.28 0 0 0 0 0 0 0 0 0 0.18 0.24 0.22 0 0 foreground (positive) background (negative) ignored GT box 0 GT box 1 ・・・ anchors matched with GT box 1, foreground matched with GT box 0, foreground multiple anchors can be assigned for one GT High recall but includes low-quality positives IoU threshold=0.71 IoU threshold=0.21 candidate anchors whose centers are close to the GT centers 

- An object detector can be decomposed into backbone, neck, dense detection head and ROI head - The core of dense detection is ground-truth sample assignment to the feature map - Assignment method varies among detectors - anchor based or point based - allow multiple anchors per GT or not - fixed or adaptive IoU threshold - Adaptive IoU thresholding improves performance of both anchor-based and anchor-free detectors Conclusion 

[1] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NIPS, 2015. [2] T.-Y. Lin, P. Dollar, R. Girshick, K. He, B. Hariharan, and S. Belongie. Feature pyramid networks for object detection. In CVPR, 2017. [3] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo and Ross Girshick, Detectron2. https://github.com/facebookresearch/detectron2, 2019. [4] Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross Girshick. Mask R-CNN. In ICCV, 2017. [5] T.-Y. Lin, P. Goyal, R. Girshick, K. He, and P. Dollar. Focal loss for dense object detection. In ICCV, 2017. [6] Mingxing Tan, Ruoming Pang, and Quoc V Le. EfficientDet: Scalable and efficient object detection. In CVPR , 2020. [7] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 779– 788, 2016. [8] Joseph Redmon and Ali Farhadi. YOLO9000: better, faster, stronger. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 7263– 7271, 2017. [9] Joseph Redmon and Ali Farhadi. YOLOv3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. [10] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934, 2020. [11] YOLOv5, https://github.com/ultralytics/yolov5 , as of version 3.0 [12] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C 12 Berg. SSD: Single shot multibox detector. In ECCV, 2016. [13] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. FCOS: Fully convolutional one-stage object detection. In ICCV, 2019. [14] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. [15] Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In CVPR, 2020. References 

Hiroto Honda’s medium blogs [H1] Digging Into Detectron 2 [part 1] : Introduction - Basic Network Architecture and Repo Structure [part 2] : Feature Pyramid Network [part 3] : Data Loader and Ground Truth Instances [part 4] : Region Proposal Network [part 5]: ROI (Box) Head [H2] Reproducing Training Performance of YOLOv3 in PyTorch [Part 0]: Introduction [Part 1]: Network Architecture and channel elements of YOLO layers [Part 2]: How to assign targets to multi-scale anchors References 

No content