Research map / Video / Video segmentation and tracking
Online-Trained Upsampler for Deep Low Complexity Video Compression
Jan P. Klopp, Keng-Chi Liu, Shao-Yi Chien, Liang-Gee Chen
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 reptile initialisation pre-training (ablation) | Online-trained upsampler · eight high-resolution datasets | BD-rate and BD-PSNR | Improved |
| add codec internal features (motion vectors, relative block positions) (ablation) | Online-trained upsampler · eight high-resolution datasets | BD-rate and BD-PSNR | Improved |
| add trigonometric position encoding (ablation) | Online-trained upsampler · eight high-resolution datasets | BD-rate and BD-PSNR | Improved |
| use only low-frequency position encodings instead of high frequencies (ablation) | Online-trained upsampler · position encoding | predictive capability | Improved |
| use position-aware upsampler instead of online-trained denoiser [16] | Online-trained upsampler · JVET A (32 frames) | BD-rate: -5.7% -> -18.4% | Improved |
| use position-aware upsampler instead of online-trained denoiser [16] | Online-trained upsampler · Netflix (32 frames) | BD-rate: -5.0% -> -27.5% | Improved |
| use position-aware upsampler instead of online-trained denoiser [16] | Online-trained upsampler · CTC A (32 frames) | BD-rate: -9.1% -> -13.6% | Improved |
| increase frame group size from 32 to 128 (ablation) | Online-trained upsampler · Netflix | BD-rate: -27.5% -> -20.6% | Hurt |
| increase frame group size from 32 to 128 (ablation) | Online-trained upsampler · JVET A | BD-rate: -18.4% -> -19.6% | Improved |
| increase frame group size from 32 to 128 (ablation) | Online-trained upsampler · JVET A UHD, QP 22 | encoding time per frame: 2.40s -> 1.11s | Improved |
| train on only the first N frames of a 32-frame group, N below half (ablation) | Online-trained upsampler · offline generalisation | BD-rate | Hurt |
| raise signalling frequency from 8 to 32 frames in zero-latency mode (ablation) | Online-trained upsampler · Netflix (zero-latency) | BD-rate: -17.2% -> -8.6% | Hurt |
Closest papers by meaning
- Neural Rate Control for Learned Video Compression (ICLR 2024 · similarity 0.85)
- Computationally-Efficient Neural Image Compression with Shallow Decoders (ICCV 2023 · similarity 0.85)
- Efficient Video Compression via Content-Adaptive Super-Resolution (ICCV 2021 · similarity 0.84)
- Few-Shot Domain Adaptation for Learned Image Compression (AAAI 2025 · similarity 0.84)
- Good, Cheap, and Fast: Overfitted Image Compression with Wasserstein Distortion (CVPR 2025 · similarity 0.83)
- C3: High-Performance and Low-Complexity Neural Compression from a Single Image or Video (CVPR 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.
