Research map / Architectures and efficiency / CNN design
Faster Neural Networks Straight from JPEG
Read the paperCheck 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 |
|---|---|---|---|
| feed DCT coefficients with receptive-field-aware layers instead of RGB pixels | ResNet-50 · ImageNet val | top-5 error %: 7.35 -> 6.98 | Helped |
| feed DCT coefficients, Late-Concat-RFA-Thinner | ResNet-50 · ImageNet val | inference FPS: 208 -> 369 | Helped |
| use YCbCr instead of RGB input (ablation) | ResNet-50 · ImageNet val | top-1 error %: 24.22 -> 24.36 | No clear effect |
| add receptive-field-aware blocks to UpSampling (ablation) | UpSampling · ImageNet val | top-1 error %: 25.07 -> 24.06 | Helped |
| add receptive-field-aware blocks to Late-Concat (ablation) | Late-Concat · ImageNet val | top-1 error %: 24.93 -> 24.08 | Helped |
| add RFA to Late-Concat (ablation) | Late-Concat · ImageNet val | inference FPS: 401 -> 267 | Hurt |
| downsample Y channel instead of upsampling chroma (ablation) | UpSampling · ImageNet val | top-1 error %: 25.07 -> 27.00 | Hurt |
| learned deconvolution instead of upsampling (ablation) | UpSampling-RFA · ImageNet val | top-5 error %: 7.14 -> 6.98 | Helped |
| thin channels of Late-Concat-RFA (ablation) | Late-Concat-RFA · ImageNet val | inference FPS: 267 -> 369 | Helped |
| remove identity blocks from baseline ResNet-50 RGB (ablation) | ResNet-50 RGB · ImageNet val | top-5 error % | Hurt |
| reduce channels of baseline ResNet-50 RGB (ablation) | ResNet-50 RGB · ImageNet val | top-5 error % | Hurt |
| learn DCT-like first-layer filters (DCT-Learn/Ortho) vs frozen DCT filters (ablation) | ResNet-50 YCbCr pixels · ImageNet val | top-5 error % | Hurt |
Closest papers by meaning
- RGB No More: Minimally-Decoded JPEG Vision Transformers (CVPR 2023 · similarity 0.88)
- CNNpack: Packing Convolutional Neural Networks in the Frequency Domain (NeurIPS 2016 · similarity 0.87)
- Doubly Convolutional Neural Networks (NeurIPS 2016 · similarity 0.86)
- P-nets: Deep Polynomial Neural Networks (CVPR 2020 · similarity 0.85)
- Curriculum By Smoothing (NeurIPS 2020 · similarity 0.85)
- JPEG Inspired Deep Learning (ICLR 2025 · similarity 0.84)
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.