Abstract
Scale variation stands out as one of the key challenges in face detection. Recent attempts have been made to cope with this issue by incorporating image/feature pyramids or adjusting anchor sampling/matching strategies. In this work, we propose a novel scale-aware progressive training mechanism to address large scale variations across faces. Inspired by curriculum learning, our method gradually learns large-to-small face instances. The preceding models learned with easier samples (i.e., large faces) can provide good initialization for succeeding learning with harder samples (i.e., small faces), ultimately deriving a better optimum of face detectors. Moreover, we propose an auxiliary anchor-free enhancement module to facilitate the learning of small faces by supplying positive anchors that may not be covered according to the criterion of IoU overlap. Such an anchor-free module is removed during inference, introducing no extra computation cost. Extensive experimental results demonstrate the superiority of our method compared to the state-of-the-arts on the standard FDDB and WIDER FACE benchmarks.
Introduction
In modern face detection, scale variation is a primary obstacle. A single image may contain faces ranging from high-resolution close-ups to tiny, blurred clusters in the background. Traditional deep learning detectors often struggle with this for two reasons: optimization complexity and the "Small Face" anchor gap. Feeding faces of all sizes into a network simultaneously makes it difficult for the model to find a stable optimum. Furthermore, most detectors rely on "anchors" (pre-defined boxes) that must overlap significantly with a ground-truth face to be considered a positive training sample. For tiny faces, even a minor misalignment causes the Intersection over Union (IoU) to drop below the threshold, leaving the model with insufficient data to learn from.
Methodology
Scale-Aware Progressive Learning
Inspired by curriculum learning, ProgressFace adopts a "large-to-small" training strategy. Rather than attempting to master all scales at once, the model begins by training on large, "easy" faces. As training progresses, it gradually incorporates smaller, "harder" faces. This progression allows the network to develop a strong "effective receptive field" and leverage context learned from larger faces to identify smaller ones later in the process.
Auxiliary Anchor-Free Enhancement Module
To bridge the anchor matching gap, the authors introduce an auxiliary branch that predicts face centers and sizes directly, similar to anchor-free methods like CenterNet. This branch identifies small faces that the standard anchor-based system misses due to low IoU. Crucially, this module acts as a "teacher" during training to help the main network learn better features; it is discarded during inference, ensuring the final model remains efficient.
Uncertainty Estimation in Localization
Real-world data often contains ambiguous boundaries due to occlusion or blur. ProgressFace addresses this by predicting a probability distribution (Gaussian) for bounding box edges rather than a single coordinate. By minimizing the Kullback–Leibler (KL) divergence, the model accounts for localization uncertainty, making it more robust to inaccurate human annotations in crowded datasets.
Experimental Evaluation
The model was evaluated on the WIDER FACE and FDDB benchmarks. The results are particularly significant for lightweight models:
- Performance Gains: Using a MobileNet-0.25 backbone, ProgressFace-Light achieved 87.9% AP on the WIDER FACE "Hard" set, outperforming RetinaFace by 9.7%.
- State-of-the-Art: The ResNet-152 variant achieved 91.8% AP on WIDER FACE Hard and a 98.7% True Positive Rate on FDDB.
- Efficiency: The model maintains high-speed performance (1.35G FLOPs for the light version) because the anchor-free module is removed after training.
Conclusion
ProgressFace demonstrates that curriculum-based progressive learning and auxiliary anchor-free supervision can significantly improve the detection of small faces without increasing inference latency. While the current fusion of anchor-based and anchor-free paradigms is limited to training, future work may explore more sophisticated inference-time integration to further push the boundaries of scale-aware detection.
Summary updated