Research map / Video / Video segmentation and tracking
Bias for Action: Video Implicit Neural Representations with Bias Modulation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ActINR, a video implicit neural representation that shares frame-INR weights across frames while using per-frame biases produced by a time-conditioned bias INR, evaluated on interpolation, super-resolution, denoising and inpainting.
Problem
Continuous video representation with implicit neural representations struggles to model large spatial or temporal motion, limiting performance on tasks like frame interpolation, slow motion, denoising and inpainting.
Approach
Treat INR biases (not weights) as the carrier of motion: weights are shared across frames as basis-function shapes, while biases are generated by a separate time-conditioned INR so that compact wavelet basis functions translate smoothly over time.
Key results
- 2x interpolation on UVG averages 31.0 dB PSNR vs 30.6 dB for H-NeRV Boost and 29.1 dB for ResFields; on DAVIS averages 22.9 dB vs 22.2 dB for ResFields.
- Video denoising on DAVIS averages 29.0 dB vs 25.5 dB for H-NeRV Boost (about 3-5 dB gain).
- Spatio-temporal super-resolution at 2x spatial with 2x time gives 34.6 dB vs 26.1 dB for H-NeRV Boost (8.5 dB gain).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ActINR instead of ResFields | Video interpolation · DAVIS 2× interpolation (average) | PSNR: 22.2 -> 22.9 | Helped |
| Use ActINR instead of H-NeRV Boost | Video interpolation · UVG 2× interpolation (average) | PSNR: 30.6 -> 31.0 | Helped |
| Use ActINR instead of H-NeRV Boost | Video denoising · DAVIS (average) | PSNR: 25.5 -> 29.0 | Helped |
| Replace bias-INR with linear interpolation (ablation) | ActINR · Toy video interpolation | PSNR: 45.8 -> 20.2 | Hurt |
| Use a bias-INR interpolator rather than assigning biases to all frames (ablation) | ActINR · Toy video interpolation | PSNR: 46.3 -> 45.8 | No clear effect |
| Use compact, oscillatory WIRE activation rather than SIREN (ablation) | ActINR · Toy moving-circle interpolation | PSNR: 30.2 -> 40.9 | Helped |
| Use WIRE rather than Gauss activation (ablation) | ActINR · Toy moving-circle interpolation | PSNR: 39.3 -> 40.9 | Helped |
Surprising findings
- Linear interpolation of biases drops test PSNR by about 24 dB relative to training, while the bias-INR nearly matches an oracle with per-frame biases.
- NeRV-based methods are reported as unable to perform space-time super-resolution beyond temporal upscaling.
Limitations
- Assumes motion stays within the designated block size; objects leaving block boundaries degrade performance and larger blocks increase parameters.
- Compression performance is slightly worse than Hybrid NeRV, though better than JPEG2000.
- Method is per-video overfitting and evaluated on a limited set of tasks and videos.
Experimental setup
Datasets: UVG DAVIS
Benchmarks: UVG DAVIS
Metrics: PSNR MS-SSIM
Baselines: ResFields FF-NeRV H-NeRV Boost DS-NeRV Phase-INR SIREN Gauss
Builds on: WIRE NeRV KiloNeRF Random Fourier Features
Closest papers by meaning
- Explaining the Implicit Neural Canvas: Connecting Pixels to Neurons by Tracing their Contributions (CVPR 2024 · similarity 0.88)
- Boosting Neural Representations for Videos with a Conditional Decoder (CVPR 2024 · similarity 0.87)
- VI^3NR: Variance Informed Initialization for Implicit Neural Representations (CVPR 2025 · similarity 0.87)
- Implicit Representations via Operator Learning (ICML 2024 · similarity 0.86)
- Inductive Gradient Adjustment for Spectral Bias in Implicit Neural Representations (ICML 2025 · similarity 0.86)
- Implicit Motion Function (CVPR 2024 · 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.