Research map / Video / Video segmentation and tracking
Diagnosing and Preventing Instabilities in Recurrent Video Processing
Thomas Tanay, Aivar Sootla, Matteo Maggioni, Puneet K. Dokania, Philip Torr, Ales Leonardis, Gregory Slabaugh
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 Hard Lipschitz constraint SRN-C-1.0-1.0 (spectral norm 1) (ablation) | VResNet-feat · Vimeo-90k denoising | PSNR_7 (dB): 35.74 -> 35.31 | Hurt |
| add Soft Lipschitz constraint SRN-C-2.0-0.1 (ablation) | VResNet-feat · Vimeo-90k denoising | PSNR_7 (dB): 35.74 -> 35.59 (-0.15 dB) | Hurt |
| use Soft instead of Hard Lipschitz constraint (alpha=2.0, beta=0.1 vs alpha=1.0, beta=1.0) (ablation) | SRN-C on VResNet-feat · Vimeo-90k denoising | PSNR_7 (dB): 35.31 -> 35.59 (+0.18dB) | Improved |
| lower alpha from 2.0 to 1.0 in SRN-C (beta=1.0) (ablation) | VResNet-feat · Vimeo-90k denoising | PSNR_7 (dB): 35.71 -> 35.31 (-0.4dB) | Hurt |
| use SRN-C instead of SRN (normalise the conv layer not the reshaped kernel) (ablation) | VResNet-feat · Vimeo-90k denoising stability | spectral norm of conv layers: around 2.5 (±0.2) for SRN -> set precisely to alpha | Improved |
| reduce alpha below 1 (SRN-C, beta=1.0) (ablation) | VResNet-feat · Vimeo-90k denoising | long-sequence stability: alpha>1 unstable, alpha<1 stable | Improved |
| reduce stable rank parameter beta below 0.1 (alpha=2.0) (ablation) | VResNet-feat · Vimeo-90k denoising | long-sequence stability: beta>0.1 unstable, beta<0.1 stable | Improved |
| add SRN-C-2.0-0.05 soft constraint | VResNet-feat · Vimeo-90k 2x super-resolution | PSNR_7 (dB): 32.58 -> 32.64 (+0.06 dB) | Improved |
| add SRN-C-1.0-1.0 hard constraint | VResNet-feat · Vimeo-90k 2x super-resolution | PSNR_7 (dB): 32.58 -> 32.57 | No clear effect |
| add feature dampening lambda=0.55 at inference (ablation) | VResNet-feat · Vimeo-90k denoising | PSNR_7 (dB): 35.74 -> 34.62 (-1.12 dB) | Hurt |
| add recurrent connection (RLSP) to feedforward VResNet multi-frame models (ablation) | VResNet · Vimeo-90k denoising | PSNR: about 0.5dB | Improved |
| make RLSP recurrence bidirectional (ablation) | VResNet-RLSP · Vimeo-90k denoising | PSNR: another 0.4dB | Improved |
Closest papers by meaning
- Stable Long-Term Recurrent Video Super-Resolution (CVPR 2022 · similarity 0.90)
- Instant Video Models: Universal Adapters for Stabilizing Image-Based Networks (NeurIPS 2025 · similarity 0.89)
- Single-Frame Regularization for Temporally Stable CNNs (CVPR 2019 · similarity 0.89)
- Learning Blind Video Temporal Consistency (ECCV 2018 · similarity 0.88)
- Stable Recurrent Models (ICLR 2019 · similarity 0.87)
- Recurrent Neural Networks With Intra-Frame Iterations for Video Deblurring (CVPR 2019 · similarity 0.87)
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.
