HiNeRV: Video Compression with Hierarchical Encoding-based Neural Representation

Learning-based video compression is currently a popular research topic, offering the potential to compete with conventional standard video codecs. In this context, Implicit Neural Representations (INRs) have previously been used to represent and compress image and video content, demonstrating relatively high decoding speed compared to other methods. However, existing INR-based methods have failed to deliver rate quality performance comparable with the state of the art in video compression. This is mainly due to the simplicity of the employed network architectures, which limit their representation capability. In this paper, we propose HiNeRV, an INR that combines light weight layers with novel hierarchical positional encodings. We employs depth-wise convolutional, MLP and interpolation layers to build the deep and wide network architecture with high capacity. HiNeRV is also a unified representation encoding videos in both frames and patches at the same time, which offers higher performance and flexibility than existing methods. We further build a video codec based on HiNeRV and a refined pipeline for training, pruning and quantization that can better preserve HiNeRV's performance during lossy model compression. The proposed method has been evaluated on both UVG and MCL-JCV datasets for video compression, demonstrating significant improvement over all existing INRs baselines and competitive performance when compared to learning-based codecs (72.3% overall bit rate saving over HNeRV and 43.4% over DCVC on the UVG dataset, measured in PSNR).

Paper

Similar papers

Peer review

Reviewer H9aP5/10 · confidence 4/52023-07-05

Summary

This paper presents HiNeRV, a new INR model for video compression based on Hierarchically-encoded Neural Representation. HiNeRV introduces a new upsampling layer that replaces conventional sub-pixel layers in existing INRs. This layer combines bilinear interpolation with hierarchical encoding, utilizing multi-resolution local feature grids. Additionally, HiNeRV incorporates depth-wise convolutional and MLP layers to construct a deep and wide network architecture with higher capacity. Sufficient experiments on several datasets and tasks have demonstrated the effectiveness of the proposed method.

Strengths

- The ideas in this paper appear to be novel (to my knowledge) and straightforward and should be easily re-implement given the paper’s descriptions. - The authors have made many technical improvements and engineering efforts to achieve competitive coding performance. - Extensive experiments have been done to demonstrate the effectiveness and compactness of the proposed method. Although HiNeRV has not been end-to-end optimized for video compression, it still significantly outperforms previous approaches with comparable bit rates.

Weaknesses

- There are several figures that may be improved for better illustration to the readers. In particular, I don't like Figure 2 since it seems too involved and without any summary in the caption. - Although the proposed method outperforms other methods in terms of encoding quality, the proposed method introduces more computational complexity, which makes encoding and decoding slower and is not conducive to real-time applications. - There are some errors in the references, for example, reference 9 should be “HNeRV: a hybrid neural representation of videos” instead of “HNeRV: neural representations for videos”. Please check your citations.

Questions

- I am confused why the performance of HNeRV will be worse than NeRV performance, and whether the experiment is correct. - Limitations or bad case discussion and analysis are expected. - Other concerns have already been mentioned in Weakness.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Yes, the authors have addressed the limitations and potential negative societal impact of their work.

Reviewer zJ4f5/10 · confidence 4/52023-07-05

Summary

This paper presents a novel method for video compression using implicit neural representation (INR), where video information is stored in the model weights. The authors explore various advanced techniques to enhance the architecture's capabilities. The primary contribution is the utilization of bilinear interpolation with hierarchical encoding for upsampling, which enables the construction of a higher-capacity network compared to using sub-pixel convolutional layers. Experimental results demonstrate that the proposed method outperforms existing INR models and is the first INR-based video codec to outperform HEVC.

Strengths

- This paper is well-organized and easy to follow. - The authors achieve state-of-the-art INR-based video compression performance, particularly surpassing HEVC for the first time. - Detailed experimental settings and comparisons are provided.

Weaknesses

- The authors do not provide the source code. - The comparison with FFNeRV, which also exhibits superior performance to HEVC, is missing. - The analysis of different interpolation methods, such as nearest neighbor and linear interpolation, is not included

Questions

- Is the architecture the same for videos with a different number of frames? If not, how is the model's size adjusted? - Please include a comparison with FFNeRV, as it also outperforms HEVC.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

Yes

Reviewer JKXU5/10 · confidence 4/52023-07-07

Summary

This paper presents a new type of implicit neural representations (INR) with hierarchical positional encoding for video compression, named HiNeRV. Four points are proposed: (1) a new upsampling layer with bilinear interpolation and hierarchical encoding of feature grids, (2) the use of depth-wise convolution and MLP layers instead of conventional convolution layers, (3) the unified representation of frame-wise and patch-wise INR by adding padding, and (4) refined training pipeline for HiNeRV model compression. The proposed method achieves significant improvement over all existing INRs baselines and competitive performance to recent neural video codecs.

Strengths

Significant rate-distortion performance as an INR-based video codec. Largely closing the gap between INR-based method and the SOTA neural video codecs.

Weaknesses

Although the proposed method achieves very significant video compression performance, the motivations behind the major contributions are not clearly discussed. Specifically, as the origin of HiNeRF, the hierarchically-encoded neural representation comes from the proposed new upsampling layer with hierarchical encoding on local feature grids. So why this design provides better performance? The paper claims that this design is similar to a convolutional layer over constant feature map. As one of the major contributions, it would be better to provide more comprehensive discussion and comparison. Moreover, according to the ablation studies, the major performance gain compared to previous methods seems to mainly come from the use of bilinear interpolation, which is a little bit incremental. Finally, it would be better to improve the paper writing. For example, as one of the contributions, the unified frame-wise and patch-wise representation is not mentioned in the abstract. And the analysis in ablation studies can be further improved.

Questions

Below are some further questions: (1) Why the proposed method performs a bit worse on MCL-JCV dataset? (2) Is it possible to provide the best/worst compression cases? In HNeRV, the authors show their model has limited performance on highly dynamic scenes. Is there a similar issue in HiNeRV? (3) Is it possible to provide the comparison of encoding complexity with recent neural video codecs? (4) How do you compress the feature grid? How much bitrate cost of the feature grid? (5) Some typos in the paper, e.g. Line 60: before quantiszation -> before quantization

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

none

Reviewer zt3G5/10 · confidence 5/52023-07-07

Summary

The authors propose a 2D implicit neural compression method that combines bilinear interpolation and hierarchical encoding. This method is based on the premise that bilinear interpolation is a better choice for neural compression than convolution operations.

Strengths

1. The proposed method distinguishes itself from other existing methods by utilizing bilinear interpolation for feature upsampling. 2. The paper investigates model compression techniques, including pruning and quantization, to further reduce the size of the model. 3. The effectiveness of the proposed modules is convincingly demonstrated through an ablation study.

Weaknesses

1. The explanation of the hierarchical coding process in section 3.4 and Fig.2 is unclear and difficult to follow, requiring further clarification. 2. It is crucial for the paper to include results on encoding and decoding time, as this information holds significant relevance in real-world scenarios. 3. The use of deep models with overlapped patches may lead to an unfair comparison with other implicit compression methods due to the increase in training pixels it implies. However, the paper lacks sufficient explanation on this matter, particularly regarding how the final reconstructed pixels are determined by overlapping blocks.

Questions

Does the training method in this paper involve training only one model on the entire dataset for inference, as mentioned in NeRV[10]? How does this encoding/decoding process apply to actual scenarios? If so, how is the compression performance on a single-sequence dataset like bunny?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

While the paper effectively improves the compression efficiency of the implicit codec, there is scope for future research in optimizing encoding time and latency.

Reviewer boCg5/10 · confidence 4/52023-07-08

Summary

This paper introduces a hierarchical encoding block for INR-based video compression. In particular, the authors replace the commonly-used sub-pixel/transpose convolution for upsampling with the bi-linear interpolation. The saved parameters are used to increase the capacity of the MLP encoders in order to achieve higher expressiveness of the model. Reportedly, this work is the first INR-based method that achieves comparable rate-distortion performance to the state-of-the-art VAE-based learned methods.

Strengths

(1) The paper is easy to follow. (2) The coding results are promising.

Weaknesses

(1) The novelty of this paper is rather now. Essentially, the authors replace the commonly-used sub-pixel/transpose convolution for upsampling with the bilinear interpolation. In doing so, the saved parameters are used to increase the number of MLP encoders. The gain appears to come mainly from a simple tweak to the network architecture. (2) The network architecture of the proposed method is not new, and is more like a trivial combination of several existing works. For example, the Grid and Stem modules are from [1] and [2]. After index transformation, the feature upsampling part is from [3]. [1] J. C. Lee, D. Rho, J. H. Ko, and E. Park. FFNeRV: Flow-guided frame-wise neural representations for videos. arXiv preprint arXiv:2212.12294, 2022. [2] J. L. Ba, J. R. Kiros, and G. E. Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016 [3] Z. Liu, H. Mao, C.-Y. Wu, C. Feichtenhofer, T. Darrell, and S. Xie. A convnet for the 2020s. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. (3) The authors should include the comparison with DCVC-HEM in the main paper. Not sure why DCVC (which is a weaker baseline) is used instead. Also, please clarify the GOP size. It seems that GOP=12 is adopted for DCVC-HEM. It was argued in several papers (including DCVC-HEM) that a larger GOP should be used. (4) The comparison with HNeRV may be unfair, because the network of HNeRV is designed to represent the whole dataset. As such, its network is quite heavy, and can actually be reduced considerably when it is used to represent only one sequence. In this sense, the per-sequence rate reported for HNeRV may be overestimated. [4] H. Chen, M. Gwilliam, S.-N. Lim, and A. Shrivastava. HNeRV: Neural representations for videos. In CVPR, 2023.

Questions

See my comments in the weaknesses section.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

The encoding time is not addressed.

Reviewer boCg2023-08-16

Comparison with DCVC-HEM

I thank the authors for putting effort into addressing my comments. Regarding the comparison with DCVC-HEM, it seems that the proposed method still encodes the entire sequence instead of 96 frames. This makes the comparison inconclusive.

Authorsrebuttal2023-08-16

We thank the reviewer for the further comment. To clarify, we conducted the experiments using full length video sequences. This ensured that the results were fairly compared and more representative. Hence both HiNeRV and DCVC-HEM were evaluated under identical conditions, using complete video sequences. We therefore consider our comparison to be fair and conclusive.

Reviewer boCg2023-08-22

Thanks for the clarification. I would change my rating to weak accept.

Reviewer JKXU2023-08-18

Thanks for the author’s detailed response and more new results. Some of my concerns have been addressed. Although it still requires a major change from the initial version, I am going to increase my rating.

Reviewer zJ4f2023-08-18

Thank you for the authors' efforts in the rebuttal. Their response has addressed my concerns, and I have increased my rating.

Reviewer zt3G2023-08-19

Improved Performance

Thank you for the author's response, it clarified some issues. In my opinion, the main contribution of this paper is indeed a relatively small but also clever improvement in terms of technique, which brings noticeable performance gains to the overall system. One major issue with this paper is its writing aspect, I believe it needs substantial revisions. The proposed hierarchical encoding does not appear to have much novelty from my perspective, although this term is frequently mentioned. Overall, there are strong advantages in terms of performance but lacking novelty and requiring further improvements in writing. I will maintain my current score.

Authorsrebuttal2023-08-21

We appreciate the additional comments from Reviewer zt3G. Please note that these issues have also been raised by other reviewers, and we have addressed them in the previous responses. Regarding novelty, please refer to our response to reviewer boCg, and for issues related to the novelty of (and motivation for) hierarchical encoding, we have clarified addressed this in our response to reviewer JKXU. For convenience, we summarize our replies and provide some additional information here. **The overall novelty** 1\. We utilized parameter efficient layers for INR-based video compression. In INR-based compression, the number of parameters is directly related to the compression performance. However, all previously published work has employed layers with high complexity, e.g. sub-pixel convolution. In contrast, in HiNeRV, we use a bilinear interpolation layer, depth-wise convolutional layers and MLPs as the network layers. Despite the simplicity of this innovation, we have clearly shown that significant gains can be obtained by adopting parameter efficient layers. 2\. We also proposed hierarchical encoding, which encodes the relative positional information. Hierarchical encoding is parameter efficient as it scales with the upsampling ratio of the layer, rather than the full spatial dimensions. By adding multiple hierarchical encodings at different layers, we can represent high resolution signals, without the need for feature grids with high storage cost. When fixing the number of parameters, hierarchical encoding provided significant improvement for highly dynamic video sequences where previous works (e.g. NeRV/HNeRV) performed poorly (0.78/1.06 dB in PSNR for the Jockey/ReadySetGo sequence). In our reply to reviewer JKXU, we mentioned that the high performance of HiNeRV with the MLP block is due to the use of the hierarchical encoding. To validate this, we performed an additional experiment and found that if we remove the hierarchical encoding from HiNeRV (with MLP blocks) then the average PSNR decreased by 2.54 dB. HiNeRV with MLP blocks, each of which only has a very lightweight convolutional layer in the stem, clearly outperformed the state-of-the-art CNN-based approach, HNeRV. The high performance of the MLP-based network is also different from the observation in previous work (the NeRV paper), which reported that MLP-based networks perform poorly in the context of the video encoding task. This further confirms the effectiveness of hierarchical encoding. We believe the use of the hierarchical encoding could also benefit MLP-based INR, but this remains for future research. 3\. While existing INR approaches for video compression are either frame- or patch-based, we introduced a unified representation. Our proposed HiNeRV, as a unified representation, can thus be utilized with either frames or overlapped patches. It yields better performance than a frame-wise representation, but also benefits from the parallelism of a patch-wise representation which can reduce the minimum memory requirement for computation. The performance improvement is confirmed in the ablation study (Section 4.3 in the paper), where we observe that HiNeRV with the unified representation obtained up to 0.49 dB PSNR improvement over the frame-wise configuration, on the UVG dataset. We also found that the improvement is more significant when the encoding time is reduced. We performed an additional experiment following the settings used in Section 4.1 with only 37 epochs, and we observed that HiNeRV with the unified representation obtained 0.93 dB improvement in average PSNR compared with the frame-wise configuration. Moreover, with our unified representation, we can perform computation using patches without sacrificing performance - an important characteristic for INR-based video compression, due to the higher signal dimensionality. For example, a recently published paper [1] had to divide the videos into patches due to memory constraints. **Writing of the paper** We appreciate the comments provided by the reviewers regarding the paper's writing. We will revise the paper in order to enhance its clarity and readability. [1] He et.al, Towards Scalable Neural Representation for Diverse Videos, CVPR 2023

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC