Research map / Video / Video segmentation and tracking
Multi-Edge Reinforced Collaborative Data Acquisition for Continuous Video Analytics by Prioritizing Quality over Quantity
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 a multi-edge collaborative reinforcement-learning policy (Transformer encoder with actor-critic decoder) that selects a small high-quality subset of video frames across edge cameras for continuous learning, cutting labeling and training data by roughly 60-70%.
Problem
Continuous learning for edge video analytics under data drift is costly because all drifted video frames must be labeled and used for retraining; frames are highly redundant across cameras and over time, and existing data acquisition methods assume a single static dataset.
Approach
An active video acquisition (AVA) policy network, a Transformer encoder with actor-critic heads, is trained with reinforcement learning across multiple edge nodes via gradient synchronization. Its reward combines task accuracy, a subset-size penalty, and diversity and representativeness rewards, so it selects frames that are informative and non-redundant across edges.
Key results
- On CORe50 classification with a budget of at most 20%, the method reaches 64.02% Top-1 versus 62.17% for the best baseline (Coreset) and 59.84% for Random.
- On VisDrone detection with a budget of at most 30%, the method reaches 56.68 AP versus 53.14 for the best baseline (MDN); auxiliary rewards improve small/medium/large object precision.
- With 20 cameras, multi-edge AVA acquires under 20% of frames while maintaining 60% training accuracy on ResNet-18.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AVA instead of Coreset for frame selection | CORe50 classification · CORe50 | Top-1 accuracy (%): 62.17±1.51 -> 64.02±0.83 | Helped |
| Use AVA instead of MDN for frame selection | Faster R-CNN · VisDrone | AP: 53.14±0.44 -> 56.68±0.60 | Helped |
| Use AVA instead of BALD for frame selection | Faster R-CNN · MOT15-KITTI | AP: 29.38±0.61 -> 31.91±0.67 | Helped |
| Add R/small auxiliary reward (ablation) | AVA with Faster R-CNN · VisDrone | AP: 56.68±0.60 -> 53.12±0.66 | Hurt |
| Add R/med auxiliary reward (ablation) | AVA with Faster R-CNN · VisDrone | AP: 56.68±0.60 -> 54.19±0.73 | Hurt |
| Add R/large auxiliary reward (ablation) | AVA with Faster R-CNN · VisDrone | AP: 56.68±0.60 -> 54.42±0.41 | Hurt |
| Add R/small auxiliary reward (ablation) | AVA with Faster R-CNN · MOT15-KITTI | AP: 31.91±0.67 -> 32.14±1.44 | Helped |
| Add R/med auxiliary reward (ablation) | AVA with Faster R-CNN · MOT15-KITTI | AP: 31.91±0.67 -> 31.43±0.81 | Hurt |
| Add R/large auxiliary reward (ablation) | AVA with Faster R-CNN · MOT15-KITTI | AP: 31.91±0.67 -> 30.86±0.92 | Hurt |
Surprising findings
- Accuracy does not grow linearly with the volume of selected training data; marginal benefit diminishes as more frames are included.
- The method tends to favor larger objects, causing a selection bias that hurts APs and APm without auxiliary size-based rewards.
Limitations
- Selection bias toward large objects, leading to weaker APs and APm in some settings.
- Requires an upper bound on the training-data budget set by the user (lambda hyperparameter).
Experimental setup
Datasets: CORe50 VisDrone MOT15
Benchmarks: CORe50 VisDrone MOT15
Metrics: Top-1 accuracy Top-3 accuracy Top-5 accuracy AP AP50 AP75 APs APm APl AR
Baselines: Random Entropy Coreset MDN BALD BADGE CDAL-CS ALFA-Mix
Builds on: Transformer A2C actor-critic Coreset BADGE BALD
Closest papers by meaning
- Towards Efficient Object Re-Identification with a Novel Cloud-Edge Collaborative Framework (AAAI 2025 · similarity 0.83)
- ReRAW: RGB-to-RAW Image Reconstruction via Stratified Sampling for Efficient Object Detection on the Edge (CVPR 2025 · similarity 0.82)
- E4: Energy-Efficient DNN Inference for Edge Video Analytics via Early Exiting and DVFS (AAAI 2025 · similarity 0.82)
- EventPillars: Pillar-based Efficient Representations for Event Data (AAAI 2025 · similarity 0.82)
- Asynchronous Collaborative Graph Representation for Frames and Events (CVPR 2025 · similarity 0.82)
- Learning Temporally Consistent Video Depth from Video Diffusion Priors (CVPR 2025 · similarity 0.81)
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.