Research map / Video / Video segmentation and tracking
Real-Time Video Inference on Edge Devices via Adaptive Model Streaming
Mehrdad Khani, Pouya Hamadanian, Arash Nasr-Esfahany, Mohammad Alizadeh
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add continual online adaptation instead of no customization | DeepLabV3-MobileNetV2 · Four video datasets | mIoU: 0.4--17.8% (8.3% on average) better | Improved |
| use continual adaptation instead of one-time customization on first 60 s (ablation) | AMS · A2D2 | mIoU: 50.78 -> 69.31 | Improved |
| one-time customization versus none | DeepLabV3-MobileNetV2 · A2D2 | mIoU: 62.05 -> 50.78 | Hurt |
| use remote teacher with tracking instead of AMS | Remote+Tracking · Outdoor Scenes | mIoU: 74.26 -> 69.05 | Hurt |
| add gradient-guided selection instead of random selection (1% params) (ablation) | AMS coordinate descent · Outdoor Scenes | mIoU diff vs full model: -5.29 -> -2.87 | Improved |
| select 5% of parameters by gradient guidance rather than updating full model (ablation) | AMS · Outdoor Scenes | mIoU / downlink bandwidth: -0.73 mIoU; 3302 -> 205 Kbps | Mixed |
| update first or last layers only instead of gradient-guided (5%) (ablation) | AMS coordinate descent · Outdoor Scenes | mIoU diff vs full model: -0.73 -> -8.98 (last), -8.37 (first) | Hurt |
| use longer training horizon (about 4 min) rather than most recent frame (Just-In-Time) (ablation) | AMS · Cityscapes, A2D2, Outdoor Scenes | downlink bandwidth for same accuracy: about 10x less bandwidth | Improved |
| remove gradient-guided selection from Just-In-Time (ablation) | Just-In-Time · Cityscapes, A2D2, Outdoor Scenes | downlink bandwidth: 150x more bandwidth than AMS | Hurt |
| use training horizon 16 s versus longer (64, 256 s) at larger update intervals (ablation) | AMS student · Driving in LA | accuracy | Hurt |
| add adaptive training rate (ATR) (ablation) | AMS multi-client · Outdoor Scenes | supported edge devices per V100 with <1% mIoU loss: 7 -> 9 | Improved |
| use Just-In-Time default threshold (75%) instead of matched accuracy (ablation) | Just-In-Time · All datasets | mIoU / bandwidth: +1.0% mIoU, 3.3x higher bandwidth | Mixed |
Closest papers by meaning
- Online Model Distillation for Efficient Video Inference (ICCV 2019 · similarity 0.86)
- Towards Real-Time Segmentation on the Edge (AAAI 2023 · similarity 0.85)
- DeDelayed: Deleting Remote Inference Delay via On-Device Correction (CVPR 2026 · similarity 0.84)
- EdgeTAM: On-Device Track Anything Model (CVPR 2025 · similarity 0.84)
- E4: Energy-Efficient DNN Inference for Edge Video Analytics via Early Exiting and DVFS (AAAI 2025 · similarity 0.84)
- Stepping Forward on the Last Mile (NeurIPS 2024 · similarity 0.83)
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.
