Abstract
Pedestrian detection benefits from deep learning technology and gains rapid development in recent years. Most detectors follow general object detection frames, i.e., default boxes and two-stage processes. Recently, anchor-free and one-stage detectors have been introduced into this area. However, their accuracies are often unsatisfactory. To enjoy the simplicity of anchor-free detectors and the accuracy of two-stage ones simultaneously, we propose several adaptations based on the Center and Scale Prediction (CSP) detector. The main contributions of our paper are: (1) We improve the robustness of CSP and make it easier to train. (2) We propose a novel method to predict width, namely compressing width. (3) We achieve the second-best performance on the CityPersons benchmark, showing that an anchor-free and one-stage detector can still have high accuracy. (4) We explore capabilities of Switchable Normalization (SN) that were not mentioned in its original paper.
Introduction
Pedestrian detection is a critical component of autonomous driving, where missing a single detection can have life-threatening consequences. Historically, researchers have relied on "anchor-based" two-stage detectors (like Faster R-CNN) because they are highly accurate. However, these models are complex and computationally expensive.
Recently, "anchor-free" one-stage detectors like Center and Scale Prediction (CSP) have emerged, offering a simpler approach by predicting the center point and height of pedestrians directly. However, the original CSP is highly sensitive to batch sizes—often failing to converge if the batch is too small or too large—and its accuracy hasn't quite matched the top-tier anchor-based models. This paper introduces Adapted Center and Scale Prediction (ACSP) to bridge this gap, making anchor-free detection both robust and highly accurate.
The ACSP Framework
The authors propose several key adaptations to the original CSP architecture to improve its stability and performance.
Switchable Normalization (SN)
Standard Batch Normalization (BN) fails when batch sizes are small because the statistics become noisy. The authors replaced all BN layers with SN, which dynamically learns to weight three different normalization methods: Batch, Layer, and Instance Normalization. This allows the model to train stably regardless of whether the batch size is 1 or 16. The most striking finding is how much normalization affects training stability; in deeper layers, the model relies more on Layer Normalization because the Batch Normalization variance becomes too noisy.
Backbone and Loss Functions
The model was upgraded from a ResNet-50 to a deeper ResNet-101 backbone. Because CSP fuses feature maps from different levels, the deeper backbone provides richer semantic information without losing the fine-grained localization details needed for small pedestrians. Additionally, while many models use "Smooth L1" loss for scale regression, the authors found that a standard Vanilla L1 loss provided better convergence and accuracy for this specific task.
Compressing Width
Most detectors use a fixed aspect ratio (typically 0.41) to determine a pedestrian's width from their height. In crowded scenes, these "standard" boxes are often too wide, leading to overlaps that confuse the Non-Maximum Suppression (NMS) process. ACSP introduces a "compressing width" formula ($w = r \cdot h$, where $r < 0.41$) to create tighter bounding boxes, improving performance in dense environments.
Experimental Results
ACSP was primarily evaluated on the CityPersons benchmark, a diverse dataset specifically annotated for pedestrian detection in urban environments.
The model achieved impressive results on the CityPersons validation set using the Log-average Miss Rate (MR) metric (lower is better):
- Reasonable Set: 9.3% MR (The second-best performance at the time of publication).
- Bare Set: 5.6% MR (Achieved the state-of-the-art result).
- Stability: Unlike the original CSP, which failed to converge on 4 out of 6 experimental batch settings, ACSP converged perfectly across all settings, from a batch size of 1 to 8.
The results demonstrate that a well-optimized, simple anchor-free detector can outperform complex models with explicit occlusion-handling sub-networks simply by being more robust and using tighter bounding box predictions.
Conclusion
ACSP significantly improves the stability and accuracy of anchor-free pedestrian detection. By replacing standard Batch Normalization with Switchable Normalization and introducing the "compressing width" technique, the framework achieves state-of-the-art results without the need for complex, specialized modules. Future work may explore "representative points" to further improve accuracy in extremely crowded scenarios.
Summary updated