Research map / Detection and segmentation / CNN features
Towards Efficient Object Re-Identification with a Novel Cloud-Edge Collaborative Framework
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 cloud-edge collaborative inference framework for object re-identification where a spatial-temporal graph network (DaCM) re-ranks edge upload sequences and allocates per-camera upload bandwidth to return the desired image faster.
Problem
Centralized ReID uploads all images to the cloud, which is impractical as camera counts grow; inference in a cloud-edge setting lacks methods for ordering uploads and allocating bandwidth so the desired image returns quickly.
Approach
A Distribution-aware Correlation Modeling network (DaCM) learns camera-to-camera spatial-temporal correlations as a graph (sinusoidal time embedding plus learnable camera embeddings with GCN-style CoMo blocks) predicting which camera an object reappears in after a time delay; the output drives cloud-side bandwidth allocation and edge-side re-ranking, and can be plugged into existing ReID methods.
Key results
- On DukeMTMC-reID, SBS + Ours (OE) reaches 96.7 R-1 and 89.8 mAP versus 95.4 and 83.0 for SBS + stReID.
- Cloud-edge DaCM with both cloud and edge inference reduces mTN from 9.56 (Pattern-CE) to 4.43 on DukeMTMC-reID.
- Ours achieves 44.74 pR-1 and 1.48 mpR, versus 0.94 pR-1 and 123.77 mpR for the visual-only centralized baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DaCM cloud-level inference (ablation) | Pattern-CE on DukeMTMC-reID · DukeMTMC-reID | mTN: 9.56 -> 5.39 | Helped |
| Add DaCM edge-level inference (ablation) | Pattern-CE on DukeMTMC-reID · DukeMTMC-reID | mTN: 9.56 -> 4.43 | Helped |
| Combine DaCM cloud- and edge-level inference (ablation) | Pattern-CE on DukeMTMC-reID · DukeMTMC-reID | mTN: 9.56 -> 4.43 | Helped |
| Add DaCM edge-level inference | SBS on DukeMTMC-reID · DukeMTMC-reID | R-1: 95.4 -> 96.7 | Helped |
| Add DaCM edge-level inference | SBS on DukeMTMC-reID · DukeMTMC-reID | mAP: 83.0 -> 89.8 | Helped |
| Replace stReID with DaCM edge-level inference | SBS on Market-1501 · Market-1501 | R-1: 96.1 -> 96.4 | Helped |
| Replace stReID with DaCM edge-level inference | TranReID on Market-1501 · Market-1501 | mAP: 89.8 -> 89.0 | Hurt |
| Use DaCM edge-level inference instead of visual similarity alone | DukeMTMC-reID precise retrieval · DukeMTMC-reID | pR-1: 0.94 -> 44.74 | Helped |
| Increase α to 100 rather than using values in the 1.0~10.0 range (ablation) | DaCM edge-level inference · DukeMTMC-reID | mTN | Helped |
| Increase β (ablation) | DaCM edge-level inference · DukeMTMC-reID | mTN | Helped |
| Increase β (ablation) | DaCM edge-level inference · DukeMTMC-reID | mAP | Helped |
| Use DaCM edge-level inference instead of stReID | TranReID on Market-1501 · Market-1501 | R-1: 96.6 -> 96.9 | Helped |
Surprising findings
- Visual-only centralized ranking achieves only 0.94 pR-1, showing that the precise desired-image protocol is very hard for standard ReID methods.
Limitations
- Requires timestamps and camera IDs, so it is unsuitable for mobile devices.
- Absolute pR-1 remains low (44.74), so the desired-image retrieval task is still challenging.
- Relies on frequency statistics combined with DaCM output for stability, indicating outlier sensitivity.
Experimental setup
Datasets: DukeMTMC-reID Market-1501
Benchmarks: DukeMTMC-reID Market-1501
Metrics: Rank-1 mAP mTN pR-K mpR
Baselines: Pattern-C Pattern-CE stReID InSTD CLIP-ReID PCB SBS TransReID
Builds on: stReID GCN PCB SBS TransReID
Closest papers by meaning
- TOP-ReID: Multi-Spectral Object Re-identification with Token Permutation (AAAI 2024 · similarity 0.84)
- Multi-Edge Reinforced Collaborative Data Acquisition for Continuous Video Analytics by Prioritizing Quality over Quantity (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)
- Alice Benchmarks: Connecting Real World Re-Identification with the Synthetic (ICLR 2024 · similarity 0.82)
- UGG-ReID: Uncertainty-Guided Graph Model for Multi-Modal Object Re-Identification (NeurIPS 2025 · similarity 0.81)
- Asynchronous Collaborative Graph Representation for Frames and Events (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.