Abstract
Detecting tiny faces in crowded images remains a significant challenge due to limited pixel information. This paper proposes the Metric Learning and Graph-Cut (MLGC) framework, which leverages semantic similarity—shared traits like scale, texture, and location—between face candidates to improve detection accuracy. By analyzing local coherence in crowds, MLGC effectively recovers low-confidence true positives and filters out false positives.
Introduction
Detecting "tiny faces" in high-density crowds is notoriously difficult. When a face occupies only a few pixels, standard detection features become unreliable. Traditional approaches often rely on upscaling or surrounding context, but these can be noisy or insufficient.
The core insight of this work is that tiny faces rarely appear in isolation; they often exist within groups that share similar semantic properties. The proposed Metric Learning and Graph-Cut (MLGC) framework exploits this "local coherence" to boost performance beyond individual box analysis.
The MLGC Framework
The MLGC framework operates as a multi-stage refinement process designed to enhance the output of high-recall face detectors.
1. Candidate Generation
The process begins with a high-recall detector to identify as many potential face candidates as possible. While this step captures most true faces, it also introduces numerous false positives.
2. Metric Learning for Similarity
To distinguish between true faces and background noise, the system learns a similarity matrix. Using a linear SVM, it calculates scores for pairs of candidate boxes based on four key cues: spatial position, scale, classification score, and deep features. This step transforms the detection problem into a binary classification task: determining if two patches are semantically similar.
3. Graph-Cut Clustering
Candidates are represented as nodes in a graph, with edges defined by the learned similarity scores. A spectral clustering algorithm (Graph-Cut) partitions the candidates into coherent groups.
4. Refinement by Voting
Within each cluster, a voting mechanism is applied. If a cluster is dominated by high-confidence face candidates, lower-scoring candidates in the same group are "promoted." This allows the model to recover "weak" true faces that would otherwise fall below a standard detection threshold.
Experimental Evaluation
The MLGC framework was evaluated on three major benchmarks: WIDER FACE, Annotated Faces in the Wild (AFW), and Pascal Faces.
- WIDER FACE: MLGC outperformed the state-of-the-art HR detector on the "Easy" subset by 0.5% on validation and 0.2% on testing sets.
- AFW & Pascal Faces: The model demonstrated superior Precision-Recall (PR) curves compared to established detectors like DPM and Headhunter.
- Similarity Metric: In standalone tests, the similarity metric achieved 79.58% average precision in identifying similar pairs, validating the use of semantic similarity as a robust signal.
Conclusion and Future Work
The MLGC framework demonstrates that semantic relationships between candidates provide a powerful prior for tiny face detection. By moving away from hard numerical thresholds and focusing on local coherence, the model can effectively "rescue" faces in crowded scenes. While the pairwise comparison approach introduces some computational complexity ($O(N^2)$), the gains in precision and recall are significant. Future work may involve optimizing the similarity matrix computation or extending this similarity-based clustering to other small-object detection domains, such as satellite imagery or medical diagnostics.
Summary updated