Research map / Label-efficient and robust learning / Recognition and long-tail
ProxyFusion: Face Feature Aggregation Through Sparse Experts
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ProxyFusion, a linear-time sparse mixture-of-experts face feature aggregation method that pools pre-computed face embeddings using learnable proxies for legacy template compatibility.
Problem
Aggregating sets of face features into a single template for long-range, low-resolution face recognition, where existing methods are quadratic-time, need intermediate feature maps or metadata, or are too costly for large probe sets.
Approach
Learnable proxies score relevance of the feature set and select the top-k sparse expert networks; each selected expert produces a set-center from the set's mean/variance and its proxy, and features are weighted by their alignment to these set-centers, using only final embeddings and giving O(N) cost.
Key results
- On BTS 3.1 face-included treatment with RetinaFace and AdaFace, ProxyFusion reaches 68.9% TAR@FAR=1e-2 versus 64.3% for CoNAN.
- With MTCNN and ArcFace on BTS 3.1, ProxyFusion reaches 31.4% TAR@FAR=1e-2 treatment, below CoNAN's 32.1%, while being linear-time.
- On DroneSURF active surveillance, ProxyFusion achieves 83.33% rank-1 versus 80.21% for CoNAN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ProxyFusion instead of CoNAN | Adaface face fusion · BTS 3.1 Face Included Treatment | TAR (%): 64.3 -> 68.9 at FAR=10−2; 49.6 -> 53.9 at FAR=10−3; 36.8 -> 40.1 at FAR=10−4 | Helped |
| use ProxyFusion instead of CoNAN | Adaface face fusion · BTS 3.1 Face Included Control | TAR (%): 96.2 -> 96.8 at FAR=10−2; 91.8 -> 92.7 at FAR=10−3; 86.1 -> 88.3 at FAR=10−4 | Helped |
| use ProxyFusion instead of CoNAN | Adaface feature fusion · DroneSURF Active, trained on DroneSURF | Rank-1 accuracy (%): 80.21 -> 83.33 | Helped |
| use ProxyFusion instead of CoNAN | Adaface feature fusion · DroneSURF Passive, trained on DroneSURF | Rank-1 accuracy (%): 13.54 -> 13.54 | No clear effect |
| train on BRS instead of DroneSURF | ProxyFusion · DroneSURF Active | Rank-1 accuracy (%): 83.33 -> 80.21 | Mixed |
| train on BRS instead of DroneSURF | ProxyFusion · DroneSURF Passive | Rank-1 accuracy (%): 13.54 -> 13.54 | No clear effect |
| use ProxyFusion instead of CAFace | Adaface fusion trained on WF4M · BTS 3.1 Face Incl. Trt., FAR=10−2 (MTCNN) | TAR (%): 51.3 -> 51.5 | Helped |
| use ProxyFusion instead of CAFace | Adaface fusion trained on WF4M · BTS 3.1 Face Incl. Trt., FAR=10−3 (MTCNN) | TAR (%): 42.0 -> 41.7 | Hurt |
| use ProxyFusion instead of CAFace | Adaface fusion trained on WF4M · BTS 3.1 Face Incl. Trt., FAR=10−4 (MTCNN) | TAR (%): 33.4 -> 33.69 | Helped |
| add proxy loss (ablation) | ProxyFusion · BTS 3.1 Face Included Treatment, selected experts bK=4 | TAR: 67.32 -> 68.93 | Helped |
| add proxy loss (ablation) | ProxyFusion · BTS 3.1 Face Included Treatment, selected experts bK=3 | TAR: 64.15 -> 65.37 | Helped |
| add proxy loss (ablation) | ProxyFusion · BTS 3.1 Face Included Treatment, selected experts bK=6 | TAR: 66.21 -> 67.03 | Helped |
Surprising findings
- Proxy loss gives only marginal gains when the number of selected experts is small.
- Increasing the number of proxies beyond about 13 leads to overfitting.
Limitations
- Relies solely on feature-set distribution statistics and ignores fine-grained intra-set relationships.
- Does not beat CoNAN on all settings (e.g., ArcFace treatment at low FAR).
Experimental setup
Datasets: BRIAR Research Set 3 WebFace 4M BTS 3.1 DroneSURF
Benchmarks: IARPA BTS 3.1 DroneSURF
Metrics: TAR@FAR rank-1 accuracy GFLOPs
Baselines: GAP NAN MCN CoNAN CAFace
Builds on: mixture-of-experts supervised contrastive loss CoNAN CAFace
Closest papers by meaning
- Cluster and Aggregate: Face Recognition with Large Probe Set (NeurIPS 2022 · similarity 0.87)
- FaceCoresetNet: Differentiable Coresets for Face Set Recognition (AAAI 2024 · similarity 0.83)
- Vec2Face: Scaling Face Dataset Generation with Loosely Constrained Vectors (ICLR 2025 · similarity 0.83)
- Proxy-FDA: Proxy-based Feature Distribution Alignment for Fine-tuning Vision Foundation Models without Forgetting (ICML 2025 · similarity 0.82)
- TopoFR: A Closer Look at Topology Alignment on Face Recognition (NeurIPS 2024 · similarity 0.82)
- HyperFace: Generating Synthetic Face Recognition Datasets by Exploring Face Embedding Hypersphere (ICLR 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.