Research map / 3D vision / Human pose estimation
ResQ: Residual Quantization for Video Perception
Davide Abati, Haitam Ben Yahia, Markus Nagel, Amirhossein Habibian
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 |
|---|---|---|---|
| quantize residuals relative to a keyframe at lower precision instead of quantizing each frame independently (ablation) | HRNet / DDRNet with PTQ · JHMDB, Cityscapes | accuracy vs BOPs trade-off | Improved |
| add dynamic per-pixel bit-width policy to ResQ (ablation) | ResQ · JHMDB pose estimation | Avg PCK and GBOPs: 94.3 PCK, 384 GBOPs -> 94.1 PCK, 176 GBOPs | Improved |
| add dynamic bit-width policy on top of static ResQ for STM (ablation) | STM VOS · DAVIS-2016 | J-Mean: 70.2 -> 82.8 | Improved |
| use ResQ instead of frame quantization at 4-bit activations for STM (ablation) | STM · DAVIS-2016 | J-Mean: 0.9 -> 70.2 | Improved |
| apply QAT instead of PTQ to ResQ (ablation) | ResQ DDRNet-23s · Cityscapes | mIoU: 66.76 -> 77.70 | Improved |
| use per-channel instead of per-tensor weight quantization (ablation) | ResQ PTQ · Cityscapes | mIoU: 73.41 -> 75.14; 66.59 -> 66.76 | Mixed |
| use pairwise residuals w.r.t. keyframe instead of recurrent frame-to-frame residuals (ablation) | ResQ · JHMDB split 1 | PCK | Improved |
| use a VideoIQ-style learned frame-level precision policy instead of residual-based policy (ablation) | Cityscapes segmentation · Cityscapes | mIoU vs BOPs | Hurt |
| apply ResQ with QAT versus the floating point model | DDRNet · Cityscapes | mIoU | Improved |
| replace 8-bit frame quantization with ResQ at W4A4/W3A3 versus Delta Distillation | HRNet-w18s · Cityscapes | mIoU and BOPs: 75.7 -> 76.9; 2.2 -> 0.9 BOPs (T) | Improved |
| quantize STM to 8 bits with frame quantization (ablation) | STM · DAVIS-2016/2017 | J-Mean: 4.9 and 3.5 drop in J-Mean | Hurt |
Closest papers by meaning
- Perception Encoder: The best visual embeddings are not at the output of the network (NeurIPS 2025 · similarity 0.87)
- PMQ-VE: Progressive Multi-Frame Quantization for Video Enhancement (NeurIPS 2025 · similarity 0.86)
- AIQViT: Architecture-Informed Post-Training Quantization for Vision Transformers (AAAI 2025 · similarity 0.86)
- Zero-Shot Quantization via Weight-Space Arithmetic (NeurIPS 2026 · similarity 0.86)
- PTQ4SAM: Post-Training Quantization for Segment Anything (CVPR 2024 · similarity 0.85)
- Quant Experts: Token-aware Adaptive Error Reconstruction with Mixture of Experts for Large Vision-Language Models Quantization (CVPR 2026 · similarity 0.85)
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.
