Abstract
While DEtection TRansformers (DETRs) are theoretically ideal for crowded scenes due to their end-to-end nature, they often suffer from high false-positive rates. This paper identifies the "forced match" trap in standard one-to-one matching as a primary cause, where ground-truth objects are matched to low-quality predictions during training. We introduce Sample Selection for Crowded Pedestrians (SSCP), a training-time refinement that filters out "unlearnable" samples and adaptively weights loss based on sample quality. SSCP achieves state-of-the-art results on CrowdHuman and Citypersons benchmarks without any inference-time overhead.
Introduction
In crowded pedestrian detection, significant object overlap poses a major challenge. Traditional CNN-based detectors rely on Non-Maximum Suppression (NMS), which often erroneously suppresses correct detections in dense crowds. DETRs address this by using a one-to-one matching strategy (Hungarian matching) to predict exactly one box per object. However, in dense environments, this strategy often forces the model to match a ground-truth pedestrian to a "poor" prediction—one that is heavily occluded or poorly localized. This "forced match" trap compels the model to learn from low-quality features, leading to a surge in false positives.
Methodology
The proposed SSCP framework addresses these issues through two primary components applied during the training phase:
Constraint-Guided Label Assignment (CGLA)
Standard DETR matching relies on a cost function based on classification and box coordinates. CGLA introduces two strict spatial constraints: a Center Constraint, requiring the predicted center to be within a specific range of the ground truth, and a Position Constraint, requiring the Intersection over Union (IoU) to exceed a threshold ($\beta$). If a matched sample fails these constraints, it is discarded as a positive and treated as a negative training sample, ensuring the model only learns from "learnable" features.
Utilizability-Aware Focal Loss (UAFL)
To handle varying sample quality, UAFL replaces standard Focal Loss with an adaptive version. It utilizes Soft Labels based on IoU, preventing the model from over-confidently predicting boxes that only partially overlap the target. Additionally, it employs an Adaptive Gamma ($\gamma$) parameter, which adjusts based on how well a sample is learned and its IoU, focusing the model's training on high-quality, under-learned samples.
Experimental Results
The authors evaluated SSCP using Iter Deformable DETR on major benchmarks. Performance is measured by Miss Rate (MR), where lower values indicate better performance.
- CrowdHuman: SSCP achieved 39.7% MR, a 1.8% improvement over the previous state-of-the-art and a 2.2% improvement over the baseline.
- Citypersons: On the "Heavy Occlusion" subset, the method reached 31.8% MR, outperforming the baseline by 0.4%.
Ablation studies confirmed that CGLA primarily reduces false positives from background noise, while UAFL refines the model's handling of occlusion. Crucially, as SSCP only modifies the training process, the resulting model maintains the same inference speed as the baseline.
Conclusion
SSCP demonstrates that the performance of DETRs in crowded scenes can be significantly enhanced by focusing on the learnability of training samples. By avoiding the "forced match" trap and adaptively weighting loss, the model becomes more robust to noise and occlusion. This approach provides a zero-cost performance boost at inference time and suggests a promising direction for other dense object detection tasks.
Summary updated