Neural Block Compression: Variable Bitrates Feature Blocks for Texture Representation
Authors: Rui Shi, Yishun Dou, Zhong Zheng, Xiangzhong Fang, Wenjun Zhang, Bingbing Ni
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate the superior texture compression quality achieved by the proposed scheme, especially at low bitrates. 5 Experiments NBC exhibits flexibility by allowing the adjustment of resolutions of multi-resolution feature blocks and the configuration of variable bitrates. This design allows it to achieve an unprecedentedly low bitrate at compression while maintaining an acceptable or comparable performance, or alternatively, to outperform other texture compression methods with an equivalent high bitrate. To validate the superior performance of NBC, we experiment extensively on diverse textures under different compression ratios. We employ bits per pixel per channel (BPPC) as a metric to quantify the compression ratio, and design multiple configurations tailored to various BPPCs. Datasets. We collect a dataset from open-source texture assets databases, ambient CG, EISKO, Kai Moisch, and Poly Haven. It contains 16 texture sets, varying from 20482 to 81922. ... Metrics. For a comprehensive quantitative evaluation, we compute PSNR, together with perceptual quality metrics LPIPS (Zhang et al. 2018), SSIM (Wang et al. 2004), and FLIP (Andersson et al. 2020). These metrics are computed across all texture sets in the dataset and all mipmap levels. |
| Researcher Affiliation | Collaboration | Rui Shi1, Yishun Dou2, Zhong Zheng2, Xiangzhong Fang1*, Wenjun Zhang1, Bingbing Ni1* 1Shanghai Jiao Tong University, Shanghai 200240, China 2Huawei, Shanghai, China |
| Pseudocode | No | The paper describes the methodology and pipeline using text and diagrams (Figure 2 and Table 1), but does not include any clearly labeled pseudocode or algorithm blocks with structured steps. |
| Open Source Code | No | Given the unavailability of the source code for NTC, we reimplement it with PyTorch, adhering to the techniques described in (Vaidyanathan et al. 2023). For MLIC++, we use the officially released code. For other methods, we utilize the publicly available command-line tools. The paper mentions using or reimplementing code for *other* methods (NTC, MLIC++, BC, ASTC, AVIF, JPEG XL), but it does not provide an explicit statement about releasing the source code for their proposed method, NBC, nor does it provide a link to a repository. |
| Open Datasets | Yes | Datasets. We collect a dataset from open-source texture assets databases, ambient CG, EISKO, Kai Moisch, and Poly Haven. It contains 16 texture sets, varying from 20482 to 81922. |
| Dataset Splits | No | The paper states that experiments are conducted on 16 texture sets and computes metrics across all texture sets and mipmap levels, but it does not specify any training/validation/test splits, percentages, or absolute sample counts for reproducibility of data partitioning. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU models, CPU types, memory amounts) used for running the experiments. It only discusses the method's efficiency in the context of hardware (GPUs, mobile devices) generally. |
| Software Dependencies | Yes | We implement NBC in PyTorch. |
| Experiment Setup | Yes | We implement NBC in PyTorch. The compressed representation of textures in NBC is a hierarchy of multi-resolution feature blocks, together with a lightweight MLP decoder. Compression. The model is trained using an auto-decoder framework (Park et al. 2019), where the multi-resolution feature blocks and the decoder are jointly optimized. We initialize each feature mipmap Zℓof the multi-resolution blocks using uniform distribution within the same quantization range as this mip-level. The training loss is measured by the L2 distance between the decompressed texture and the ground truth texture. We employ Adam optimizer (Kingma and Ba 2014) with an initial learning rate of 0.01 for multiresolution feature blocks and 0.005 for the MLP decoder. We train the model for 20k iterations and apply cosine annealing (Loshchilov and Hutter 2016) to facilitate learning rate reduction, gradually decreasing the learning rate to 0 at the end of the training process. At training time, we simulate quantization by introducing additive uniform noise to feature vectors. While within the final 10% of training iterations, explicit quantization of the feature vectors is performed to further optimize the MLP decoder to adapt to quantized features. At this stage, feature blocks are frozen, and optimization focuses on the decoder. |