Introduction
In two-stage object detection frameworks like Faster R-CNN, the model generates thousands of candidate regions (proposals). Since the vast majority of these are background (negatives), effective training requires a sampling strategy to balance positive and negative classes. Traditional heuristics, such as IoU-balanced sampling, focus exclusively on the overlap between a proposal and the ground truth. However, this approach overlooks a critical factor: the overlap between the proposals themselves.
The Inter-Proposal Overlap Problem
The authors identify that ignoring inter-proposal overlap creates a discrepancy between training objectives and inference-time processing, specifically regarding Non-Maximum Suppression (NMS). NMS removes redundant boxes by keeping only the highest-confidence detection among overlapping candidates. If a "hard negative" (a box that doesn't fit the object well) is assigned a higher confidence score than a "positive" box it overlaps with, NMS will mistakenly suppress the correct detection.
By failing to penalize high scores for these confusing, overlapping negatives during training, standard models remain prone to localization errors and jittery outputs. Empirical "upper bound" tests show that perfectly classifying negatives that overlap with positives yields a potential AP of 49.5, compared to only 44.2 when focusing solely on ground-truth overlap.
The Overlap Sampler Method
To address this, the Overlap Sampler shifts the focus toward inter-proposal relationships through two primary modules:
- Negative Overlap Sampler: This module prioritizes negatives with high overlap with positive proposals ($IoU^{p,n}$). It utilizes Soft Probability Sampling, where the range of inter-proposal overlaps is divided into $K$ bins. Negatives are sampled uniformly from these bins, ensuring the model is exposed to "hard" negatives physically close to actual objects.
- Positive Overlap Sampler: This refines the handling of positive boxes using a Loss Reweighting Scheme. Positives with the highest ground-truth overlap are assigned a higher weight ($\epsilon$), encouraging the model to prioritize high-quality "prime" examples.
Unlike loss-based samplers like OHEM, the Overlap Sampler relies on efficient geometric IoU calculations, providing the benefits of hard-negative mining without significant computational overhead during training or inference.
Experimental Results
Evaluated on the MS COCO benchmark, the Overlap Sampler demonstrated consistent improvements across various backbones:
- Faster R-CNN: Achieved a 1.5 AP increase (37.2 to 38.7) with a ResNet-50 backbone.
- Mask R-CNN: Improved mask AP by 0.8 points, indicating benefits for instance segmentation.
- State-of-the-Art Comparison: Using a ResNext-101 backbone, the method reached 42.5 AP on COCO test-dev, outperforming sophisticated samplers like Libra R-CNN and PISA.
The gains were particularly pronounced in high-precision metrics (e.g., $AP_{75}$), suggesting superior fine-tuning of localization.
Conclusion
The Overlap Sampler provides a simple yet effective strategy for training region-based detectors by aligning the training signal with the challenges of NMS. While it introduces new hyperparameters that may require tuning for specific datasets, its efficiency and consistent performance gains make it a compelling alternative to standard sampling heuristics. Future work may explore its application in one-stage detectors or in combination with advanced architectural features.
Summary updated