This document discusses and compares different methods for deep learning object detection, including region proposal-based methods like R-CNN, Fast R-CNN, Faster R-CNN, and Mask R-CNN as well as single shot methods like YOLO, YOLOv2, and SSD. Region proposal-based methods tend to have higher accuracy but are slower, while single shot methods are faster but less accurate. Newer methods like Faster R-CNN, R-FCN, YOLOv2, and SSD have improved speed and accuracy over earlier approaches.
1. Deep learning for object
detection
Wenjing Chen
*Created in March 2017, might be outdated the time you read.
Slide credit: CS231n
2. Outline
1. Introduction
2. Common methods
Region proposal based methods
R-CNN, Fast R-CNN, Faster R-CNN, R-FCN, Mask R-CNN
Single shot based methods
YOLO, YOLOv2, SSD
1. Comparison
4. Region based methods - R-CNN
Girshick, Ross, et al. "Rich feature hierarchies for accurate object detection and semantic segmentation." Proceedings of the IEEE conference on computer
vision and pattern recognition. 2014.
5. Region based methods - Fast R-CNN
Girshick, Ross. "Fast r-cnn." Proceedings of the IEEE International Conference on Computer Vision. 2015.
6. Region based methods - Faster R-CNN
Ren, Shaoqing, et al. "Faster r-cnn: Towards real-time object detection with region proposal networks." Advances in neural information processing systems.
2015.
8. Region based methods - R-FCN
Li, Yi, Kaiming He, and Jian Sun. "R-fcn: Object detection via region-based fully convolutional networks." Advances in Neural Information Processing Systems.
2016.
Average
pooling
9. Region based methods - Mask R-CNN
He, Kaiming, et al. "Mask R-CNN." arXiv preprint arXiv:1703.06870 (2017).
Object instance segmentation:
Extend Faster R-CNN by adding a
branch for predicting segmentation
masks on each RoI
Running at 5 fps
Without tricks, outperforms all existing,
single-model entries on every task in
all three tracks of the COCO suite of
challenges, including instance
segmentation, bounding-box object
detection, and person keypoint
detection !!!
10. Single shot based method - YOLO
Redmon, Joseph, et al. "You only look once: Unified, real-time object detection." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition
2016.
1. Resize input image to 448*448.
1. Run a single convolutional network.
Predicts B bounding boxes (4 coordinates + confidence) and
C class probabilities for S*S grids, encoded as an
S*S*(B*5+C) tensor.
1. Non-maximum suppression.
S*S*B bounding boxes per image and C class probabilities
for each box.
11. Single shot based method - YOLOv2
Redmon, Joseph, and Ali Farhadi. "YOLO9000: Better, Faster, Stronger." arXiv preprint arXiv:1612.08242 (2016).
YOLO problem:
1. Significant number of localization errors.
2. Low recall compared to region proposal based methods.
Improvements:
12. Single shot based method - SSD
Liu, Wei, et al. "SSD: Single shot multibox detector." European Conference on Computer Vision. Springer International Publishing, 2016.
Improvements:
1. Use a small convolutional filter to predict object categories and offsets in bounding box
locations
2. Use multiple layers for prediction at different scales.
15. Comparison
Speed
single shot > region based
Accuracy
region based > single shot
Complexity
YOLO < SSD ≤ Faster R-CNN < R-FCN < YOLOv2(?)
Hinweis der Redaktion
Batch normalization. 2% more in mAP.
High resolution classifier. 4% more in mAP.
Convolutional with anchor boxes. 69.5 mAP 81% recall to 69.2 mAP 88% recall.
Dimension clusters. Better anchor boxes priors. 60.9% to 67.2% in Avg IOU.
Direct location prediction. Solve model instability.
Fine-Grained features. 1% more in mAP.
Multi-scale training.