Abstract
Detecting pedestrians in crowded environments is hindered by two primary issues: feature contamination from occluding objects and the statistical imbalance of highly occluded samples. This paper introduces the Dynamic Instance Noise Filter (DINF), which treats occlusion as a signal denoising task. DINF utilizes instance-specific features to generate dynamic kernels that transform Region of Interest (RoI) features into a domain where noise can be removed via a Soft Thresholding Module (STM). To further enhance performance, an Iterative Noise Filter Network (INFN) and an IoU-Focal Factor (IFF) are proposed. Experimental results on the CrowdHuman and CityPersons datasets demonstrate that DINF significantly reduces miss rates and sets a new state-of-the-art for box-based detectors.
The Challenge of Occlusion in Crowds
In crowded environments, pedestrians are rarely seen in full. Standard RCNN-based detectors extract features by cropping rectangular "Regions of Interest" (RoIs). When a person is occluded by a pole or another pedestrian, a large portion of that rectangle contains "noise"—pixels belonging to the background or other objects.
This creates two major hurdles:
- Feature Contamination: The signal-to-noise ratio (SNR) of the pedestrian's feature vector drops, making it hard for the model to distinguish the target from the occluder.
- Sample Imbalance: In datasets like CrowdHuman, most pedestrians have low-to-moderate overlap. As shown in the overlap statistics, cases of extreme occlusion (the "hard" samples) are statistically outnumbered. Consequently, the model’s loss function is often dominated by "easy" cases, leading to poor performance on the most challenging occlusions.
Methodology
1. Dynamic Instance Noise Filter (DINF)
Inspired by classical wavelet denoising, DINF treats occlusion as a signal processing problem. Unlike static filters, DINF uses the instance's own feature vector to generate custom $1 \times 1$ convolutional kernels. These kernels transform RoI features into a domain where noise is represented by near-zero values.
A Soft Thresholding Module (STM) then applies a learnable threshold ($\tau$) to zero out values in the range $[-\tau, \tau]$ and shift other values toward zero. This explicitly removes low-amplitude noise while preserving the strong signal of the target pedestrian.
2. Iterative Noise Filter Network (INFN)
The authors found that denoising once isn't always enough. They stack multiple DINF modules in an iterative structure. To prevent the loss of valuable signal over multiple steps, they use dense connections, adding the output of previous iterations to the input of the next.
3. IoU-Focal Factor (IFF)
To solve the imbalance issue, the authors adapted the concept of Focal Loss to bounding box regression. The IFF modulates the loss based on the IoU (Intersection over Union) of the prediction. If a box is already well-regressed (high IoU), its contribution to the loss is down-weighted. This forces the optimizer to pay more attention to the "hard-regressed" samples—the highly occluded pedestrians.
Experimental Results
The researchers evaluated their methods on two primary benchmarks: CrowdHuman (high overlap) and CityPersons (focusing on the "Heavily Overlapped" subset).
The proposed methods significantly improved the Faster R-CNN baseline on the CrowdHuman dataset:
- AP (Average Precision): Increased from 87.36% to 89.21%.
- MR⁻² (Miss Rate): Dropped from 42.24% to 41.09%.
- JI (Jaccard Index): Improved from 79.67% to 81.28%.
When integrated into the more advanced CrowdDet detector, the system achieved a state-of-the-art 40.35% MR⁻², outperforming all other box-based methods. On CityPersons, the model improved the Miss Rate on heavily occluded pedestrians by over 1.3%.
Conclusion
DINF demonstrates that treating occlusion as a signal denoising problem is a powerful paradigm for pedestrian detection. By generating kernels on-the-fly, the model adapts its "cleaning" strategy to specific noise patterns. While the iterative approach adds some inference overhead, the gains in precision and miss rate—especially in crowded scenes—mark a significant advancement for two-stage detectors. Future work may explore applying these dynamic denoising principles to one-stage or Transformer-based architectures.
Summary updated