Research map / Video / Video segmentation and tracking
DeltaCNN: End-to-End CNN Inference of Sparse Frame Differences in Videos
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| propagate sparse frame updates end-to-end with GPU tile skipping | HRNet · Human3.6M Subject S11, Jetson Nano | speedup: 1.0 -> 6.7 | Helped |
| propagate sparse frame updates end-to-end with update truncation | EfficientDet-d1 · MOT16, Jetson Nano | speedup: 1.0 -> 5.6 | Helped |
| use sparse inference rather than dense cuDNN inference | EfficientDet-d1 · MOT16 | AP@0.5: 63.9% -> 64.0% | No clear effect |
| use sparse inference rather than dense cuDNN inference | EfficientDet-d1 · WildTrack | AP@0.5: 72.1% -> 71.6% | Hurt |
| use a larger d1 model instead of d0 | EfficientDet on MOT16 · MOT16 | AP@0.5: 55.6% -> 64.0% | Helped |
| use hybrid tile processing rather than per-tile sparse processing (ablation) | 3x3 convolution, 6x6 tile · GTX 1050, s=90% | runtime: 16.5 -> 15.8 | Helped |
| use hybrid tile processing rather than per-tile sparse processing (ablation) | 3x3 convolution, 6x6 tile · GTX 1050, s=99% | runtime: 7.7 -> 5.0 | Helped |
| use per-pixel sparse processing instead of hybrid tile processing (ablation) | 3x3 convolution, 6x6 tile · GTX 1050, s=90% | runtime: 15.8 -> 44.3 | Hurt |
| use per-pixel sparse processing instead of hybrid tile processing (ablation) | 3x3 convolution, 6x6 tile · GTX 1050, s=99% | runtime: 5.0 -> 20.4 | Hurt |
| train truncation thresholds jointly rather than auto-tuning them one-by-one (ablation) | DeltaCNN · threshold tuning experiments | threshold quality | No clear effect |
| train truncation thresholds jointly rather than auto-tuning them one-by-one (ablation) | DeltaCNN · threshold tuning experiments | training time | Hurt |
| use threshold 0.5 with 7-pixel input-mask dilation instead of threshold 0.15 without dilation (ablation) | DeltaCNN on MOT16 · MOT16 | speed: 8% | Helped |
Closest papers by meaning
- MEET: Towards Memory-Efficient Temporal Sparse Deep Neural Networks (CVPR 2025 · similarity 0.89)
- SBNet: Sparse Blocks Network for Fast Inference (CVPR 2018 · similarity 0.88)
- SeerNet: Predicting Convolutional Neural Network Feature-Map Sparsity Through Low-Bit Quantization (CVPR 2019 · similarity 0.87)
- Fast Sparse ConvNets (CVPR 2020 · similarity 0.87)
- Designing BERT for Convolutional Networks: Sparse and Hierarchical Masked Modeling (ICLR 2023 · similarity 0.87)
- CoreMatching: A Co-adaptive Sparse Inference Framework with Token and Neuron Pruning for Comprehensive Acceleration of Vision-Language Models (ICML 2025 · similarity 0.86)
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.