# Single-frame Regularization for Temporally Stable CNNs

arXiv: https://arxiv.org/abs/1902.10424 · HTML: https://mlsanity.com/paper/1902.10424/

## Key facts

- Cited by 21 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 27 Feb 2019.

## Built on by (most cited first, 21 of 21)

- [Video Demoireing with Relation-Based Temporal Consistency](https://mlsanity.com/paper/2204.02957/) (2022)
- [Temporally Consistent Video Colorization with Deep Feature Propagation and Self-regularization Learning](https://mlsanity.com/paper/2110.04562/) (2021)
- [Internal Video Inpainting by Implicit Long-range Propagation](https://mlsanity.com/paper/2108.01912/) (2021)
- [Deep Video Prior for Video Consistency and Propagation](https://mlsanity.com/paper/2201.11632/) (2022)
- [Learning Temporally and Semantically Consistent Unpaired Video-to-video Translation Through Pseudo-Supervision From Synthetic Optical Flow](https://mlsanity.com/paper/2201.05723/) (2022)
- [Low Light Video Enhancement by Learning on Static Videos with Cross-Frame Attention](https://mlsanity.com/paper/2210.04290/) (2022)
- [Object-Centric Temporal Consistency via Conditional Autoregressive Inductive Biases](https://mlsanity.com/paper/2410.15728/) (2024)
- [Temporally stable video segmentation without video annotations](https://mlsanity.com/paper/2110.08893/) (2021)
- [SALVE: Self-supervised Adaptive Low-light Video Enhancement](https://mlsanity.com/paper/2212.11484/) (2022)
- [Universal Bovine Identification via Depth Data and Deep Metric Learning](https://mlsanity.com/paper/2404.00172/) (2024)
- [HDRVideo-GAN: Deep Generative HDR Video Reconstruction](https://mlsanity.com/paper/2110.11795/) (2021)
- [DeepHS-HDRVideo: Deep High Speed High Dynamic Range Video Reconstruction](https://mlsanity.com/paper/2210.04429/) (2022)
- [UnDIVE: Generalized Underwater Video Enhancement Using Generative Priors](https://mlsanity.com/paper/2411.05886/) (2024)
- [Unsupervised HDR Imaging: What Can Be Learned from a Single 8-bit Video?](https://mlsanity.com/paper/2202.05522/) (2022)
- [Joint tone mapping and denoising of thermal infrared images via multi-scale Retinex and multi-task learning](https://mlsanity.com/paper/2305.00691/) (2023)
- [Online Overexposed Pixels Hallucination in Videos with Adaptive Reference Frame Selection](https://mlsanity.com/paper/2308.15462/) (2023)
- [Toward Accurate and Temporally Consistent Video Restoration from Raw Data](https://mlsanity.com/paper/2312.16247/) (2023)
- [Towards stable training of parallel continual learning](https://mlsanity.com/paper/2407.08214/) (2024)
- [Adapting MIMO video restoration networks to low latency constraints](https://mlsanity.com/paper/2408.12439/) (2024)
- [WaterWave: Bridging Underwater Image Enhancement into Video Streams via Wavelet-based Temporal Consistency Field](https://mlsanity.com/paper/2512.05492/) (2025)
- [Generating HDR Video from SDR Video](https://mlsanity.com/paper/2605.14703/) (2026)

## Abstract (arXiv)

Convolutional neural networks (CNNs) can model complicated non-linear relations between images. However, they are notoriously sensitive to small changes in the input. Most CNNs trained to describe image-to-image mappings generate temporally unstable results when applied to video sequences, leading to flickering artifacts and other inconsistencies over time. In order to use CNNs for video material, previous methods have relied on estimating dense frame-to-frame motion information (optical flow) in the training and/or the inference phase, or by exploring recurrent learning structures. We take a different approach to the problem, posing temporal stability as a regularization of the cost function. The regularization is formulated to account for different types of motion that can occur between frames, so that temporally stable CNNs can be trained without the need for video material or expensive motion estimation. The training can be performed as a fine-tuning operation, without architectural modifications of the CNN. Our evaluation shows that the training strategy leads to large improvements in temporal smoothness. Moreover, for small datasets the regularization can help in boosting the generalization performance to a much larger extent than what is possible with na"ive augmentation strategies.