Research map / Video / Video segmentation and tracking
Activity-Biometrics: Person Identification from Daily Activities
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 ABNet, a 3D-CNN person-identification framework for RGB videos of daily activities that disentangles biometric from appearance features via silhouette-based distillation and elastic-distortion bias learning, plus joint activity learning, evaluated on five activity-derived datasets.
Problem
Person identification from RGB videos of daily activities (not just walking), where appearance biases such as clothing color and background confound biometric feature learning and faces are blurred.
Approach
Disentangles biometric and appearance features by distilling from a silhouette-trained bias-less teacher (GaitGL) and learning appearance bias via contrastive loss against elastic-distorted videos that change identity cues while preserving appearance; jointly trains an activity head whose features are concatenated with biometric features at inference as an activity prior.
Key results
- ABNet achieves 78.76 rank-1 and 40.31 mAP on NTU RGB-AB (View+, same activity), outperforming the best SOTA Video-CAL at 75.49 / 39.86.
- On ACC-MM1-Activities ABNet reaches 80.43 rank-1 vs 77.48 for Video-CAL.
- On BRIAR-BGC3 ABNet achieves 34.38 rank-1 vs 30.57 for Image-CAL, about 4% higher.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ABNet instead of Video-CAL | Activity-based person identification · NTU RGB-AB, same activity, View+ | Rank 1 accuracy: 75.49 -> 78.76 | Helped |
| Use ABNet instead of Video-CAL | Activity-based person identification · NTU RGB-AB, same activity, View+ | mAP: 39.86 -> 40.31 | Helped |
| Use ABNet instead of PSTR | Activity-based person identification · PKU MMD-AB, same activity, View+ | Rank 1 accuracy: 84.33 -> 86.83 | Helped |
| Use ABNet instead of PSTA | Activity-based person identification · PKU MMD-AB, same activity, View+ | mAP: 50.42 -> 57.31 | Helped |
| Use ABNet instead of Video-CAL | Activity-based person identification · Charades-AB, same activity | Rank 1 accuracy: 43.91 -> 45.84 | Helped |
| Use ABNet instead of PSTA | Activity-based person identification · Charades-AB, same activity | mAP: 28.32 -> 31.58 | Helped |
| Use ABNet instead of Video-CAL | Activity-based person identification · ACC-MM1-Activities, same activity | Rank 1 accuracy: 77.48 -> 80.43 | Helped |
| Use ABNet instead of PSTA | Activity-based person identification · ACC-MM1-Activities, same activity | mAP: 50.31 -> 52.71 | Helped |
| Use ABNet instead of Image-CAL | Person identification · BRIAR-BGC3 | Rank 1 accuracy: 30.57 -> 34.38 | Helped |
| Use ABNet instead of Image-CAL | Person identification · BRIAR-BGC3 | mAP: 17.44 -> 18.78 | Helped |
| Add bias-less distillation (ablation) | RGB person-identification baseline · NTU RGB-AB, same activity, View+ | Rank 1 accuracy: 64.23 -> 69.31 | Helped |
| Add activity prior (ablation) | RGB person-identification baseline · NTU RGB-AB, same activity, View+ | Rank 1 accuracy: 64.23 -> 69.43 | Helped |
Surprising findings
- Removing face blurring yields only a minimal performance gain (79.24 vs 78.76 rank-1 on NTU RGB-AB), suggesting limited reliance on facial cues.
- Image-based CAL outperforms many video-based methods on several datasets, e.g., 73.79 vs 67.41-72.98 rank-1 on NTU RGB-AB.
Limitations
- Performance drops notably in cross-activity and cross-view settings, e.g., ACC-MM1-Activities falls from 80.43 to 68.31 rank-1.
- Relies on silhouette extraction (Mask2Former) and a silhouette-trained teacher during training, adding training complexity.
- Evaluation limited to derived datasets from activity recognition benchmarks with few activity classes in some cases.
Experimental setup
Datasets: NTU RGB+D PKU-MMD Charades ACC-MM1-Activities BRIAR-BGC3
Benchmarks: NTU RGB-AB PKU MMD-AB Charades-AB ACC-MM1-Activities BRIAR-BGC3
Metrics: rank 1 accuracy rank 5 accuracy mAP TAR @ 0.1% FAR
Baselines: CAL PSTR SCNet AIM TSF VKD BiCnet-TKS STMN PSTA SINet Video-CAL GaitGL
Builds on: GaitGL ResNet3D Mask2Former Video-CAL knowledge distillation
Closest papers by meaning
- HARDVS: Revisiting Human Activity Recognition with Dynamic Vision Sensors (AAAI 2024 · similarity 0.82)
- Recognizing Activities of Daily Living With a Wrist-Mounted Camera (CVPR 2016 · similarity 0.82)
- Exploring More from Multiple Gait Modalities for Human Identification (AAAI 2025 · similarity 0.81)
- Generalizable Sensor-Based Activity Recognition via Categorical Concept Invariant Learning (AAAI 2025 · similarity 0.81)
- DIFFER: Disentangling Identity Features via Semantic Cues for Clothes-Changing Person Re-ID (CVPR 2025 · similarity 0.81)
- PairDETR : Joint Detection and Association of Human Bodies and Faces (CVPR 2024 · similarity 0.80)
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.