Boosting Neural Representations for Videos with a Conditional Decoder

TL;DR

This paper introduces a universal boosting framework for current implicit video representation approaches, utilizing a conditional decoder with a temporal-aware affine transform module, which uses the frame index as a prior condition to effectively align intermediate features with target frames.

Abstract

Implicit neural representations (INRs) have emerged as a promising approach for video storage and processing, showing remarkable versatility across various video tasks. However, existing methods often fail to fully leverage their representation capabilities, primarily due to inadequate alignment of intermediate features during target frame decoding. This paper introduces a universal boosting framework for current implicit video representation approaches. Specifically, we utilize a conditional decoder with a temporal-aware affine transform module, which uses the frame index as a prior condition to effectively align intermediate features with target frames. Besides, we introduce a sinusoidal NeRV-like block to generate diverse intermediate features and achieve a more balanced parameter distribution, thereby enhancing the model's capacity. With a high-frequency information-preserving reconstruction loss, our approach successfully boosts multiple baseline INRs in the reconstruction quality and convergence speed for video regression, and exhibits superior inpainting and interpolation results. Further, we integrate a consistent entropy minimization technique and develop video codecs based on these boosted INRs. Experiments on the UVG dataset confirm that our enhanced codecs significantly outperform baseline INRs and offer competitive rate-distortion performance compared to traditional and learning-based codecs. Code is available at https://github.com/Xinjie-Q/Boosting-NeRV.

Figures (12)

• Figure 1: Video regression with different encoding time under 1.5M model size (left) and video compression with different model sizes(right). Our boosted methods achieve significantly better performance than the corresponding baselines.
• Figure 2: Our proposed HNeRV-Boost framework with the conditional decoder. The content-relevant embedding $\boldsymbol{y}_t$ expands its channel dimensions in stage 1 and upsamples in stages 2 to 6. The final three stages stack two SNeRV blocks with small kernel sizes to get fewer parameters, where the former upsamples features and the latter refines the upsampled features.
• Figure 3: (Left) Illustration of the temporal-aware affine transform layer and residual block. The TAT layer takes the temporal embeddings $\boldsymbol{z}_t$ to produce channel-wise scaling and shifting parameters $\boldsymbol{\gamma}_t$ and $\boldsymbol{\beta}_t$. As a result, the affine transformation is performed to the intermediate features of the previous layer. (Right) The architecture of the sinusoidal NeRV-like block. When the stride $s$ of the convolutional layer is larger than 1, it includes a pixelshuffle layer.
• Figure 4: Visual comparisons of intermediate features from different activation functions in the HNeRV-Boost model. We select the first 40 channel features from the last NeRV-like block on the first frame generation of the Bunny video.
• Figure 5: Rate-distortion curves of our boosted approaches and different baselines on the UVG dataset in PSNR and MS-SSIM. PQE denotes the three-step compression pipeline of NeRV.