Source-linked AI summary
An Implementation of Faster RCNN with Study for Region Sampling
Xinlei Chen, Abhinav Gupta
TL;DR
The paper adapts Faster RCNN’s joint-training scheme to TensorFlow and studies how region-proposal sampling affects region classification. Across VOC 2007 and COCO 2014, biased sampling toward smaller regions outperforms uniform sampling, while sufficiently long training narrows differences among biased schemes.
Problem
The study asks whether NMS-based region sampling is necessary during Faster RCNN joint training, given that NMS may bias training toward smaller regions.
Method
The paper implements Faster RCNN in TensorFlow, evaluates pipeline simplifications, and compares uniform, NMS-based, and other biased region-sampling schemes on VOC 2007 and COCO.
Results
Biased sampling generally outperforms uniform sampling; on COCO, 790k NMS iterations reach 28.3 AP on minival with TOP testing, while longer training narrows the gap between NMS and POW.
Takeaways & Limitations
Keeping small region proposals improves small-object performance, and sufficiently converged biased sampling can approach NMS-based performance.
Takeaways & Limitations
COCO results remain inconclusive because 490k iterations are insufficient for full convergence and longer-training experiments are needed.
Abstract
from arXiv · showhide
We adapted the join-training scheme of Faster RCNN framework from Caffe to TensorFlow as a baseline implementation for object detection. Our code is made publicly available. This report documents the simplifications made to the original pipeline, with justifications from ablation analysis on both PASCAL VOC 2007 and COCO 2014. We further investigated the role of non-maximal suppression (NMS) in selecting regions-of-interest (RoIs) for region classification, and found that a biased sampling toward small regions helps performance and can achieve on-par mAP to NMS-based sampling when converged sufficiently.
1. Baseline Faster RCNN with Simplification
The TensorFlow baseline simplifies Faster RCNN training and preprocessing while remaining broadly comparable to the original implementation, with particular gains for small objects.
- The implementation adapts Faster RCNN’s join-training scheme from Caffe to TensorFlow as a publicly available baseline.
- Crop-resize pooling replaces RoI pooling by resizing crops to 14 × 14 before max-pooling to 7 × 7.
- Training samples 256 regions from one image per pass instead of aggregating 128 regions from two images.The region proposal network still uses the default 256 regions.
- Removing proposals smaller than 16 pixels is redundant and hurts performance, especially for small objects.
- On VOC 2007, the TensorFlow implementation is generally on par with Caffe, while crop-resize pooling has a slight advantage over RoI pooling.
- On COCO, the baseline improves small-object AP by 4% and AR by 5%; 256 sampled regions provide a good trade-off, while larger values risk over-fitting.
2. A Study of Region Sampling
The study compares NMS-based and alternative RoI sampling schemes for Faster RCNN training and testing. It finds that sampling biased toward smaller regions can match or exceed NMS-based performance when training converges sufficiently, while TOP testing benefits from evaluating more proposals.
- Baseline NMS sampling: The original pipeline selects top-ranked proposals, applies NMS at overlap ratio 0.7, and samples RoIs for region classification.Training uses K = 12000 and k = 2000 before sampling R regions; testing uses K = 6000 and k = 300.
- Testing without NMS: NMS can be removed during testing by directly selecting top K proposals, trading increased RoI evaluation for recall and speed considerations.The study sets K = 5000 for TOP testing because mAP generally increases as K grows.
- Alternative sampling schemes: Alternative training schemes include ALL, which uses all top K regions, PRE, which matches a pretrained NMS model’s sampling ratio, and POW, which uses r(s) = s^-γ with γ = 1.PRE depends on an NMS-trained model, whereas POW does not require NMS and biases sampling according to region scale.
- VOC 2007 results: On VOC 2007, biased schemes achieve around 71% mAP, and NMS-trained models tested with TOP outperform NMS testing when K is sufficiently large.The comparison covers train/test combinations in Table 3, with NMS as the baseline and ALL, PRE, POW, and TOP as non-NMS schemes.
- COCO 2014 results: On COCO 2014, biased sampling generally outperforms uniform sampling; NMS reaches similar performance to PRE/POW later, and 790k NMS iterations achieve AP 28.3 with TOP testing.The gap between NMS and POW narrows from 1.7 at 490k iterations to 1.4 at 790k, so longer training is needed for a more conclusive comparison.