Efficient LiDAR Reflectance Compression via Scanning Serialization
Authors: Jiahao Zhu, Kang You, Dandan Ding, Zhan Ma
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that Ser Li C attains over 2 volume reduction against the original reflectance data, outperforming the state-of-the-art method by up to 22% reduction of compressed bits while using only 2% of its parameters. Moreover, a lightweight version of Ser Li C achieves 10 fps (frames per second) with just 111K parameters, which is attractive for real-world applications. |
| Researcher Affiliation | Academia | 1School of Information Science and Technology, Hangzhou Normal University, Hangzhou, China 2School of Electronic Science and Engineering, Nanjing University, Nanjing, China. |
| Pseudocode | No | The paper describes methods like 'Serialization' and 'Contextual Construction' in section 3 with detailed steps and equations, and Figure 3 shows diagrams of 'Mamba Block' and 'Attention Block'. However, it does not contain a dedicated section or figure explicitly labeled as 'Pseudocode' or 'Algorithm' with structured, code-like steps for the overall methodology. |
| Open Source Code | No | The paper mentions 'https://github.com/MPEGGroup/ mpeg-pcc-tmc13' in reference to G-PCC, which is a method used for comparison, not the authors' own implementation of Ser Li C. There is no explicit statement or link provided for the open-sourcing of Ser Li C. |
| Open Datasets | Yes | We conducted experiments on well-known Li DAR datasets, including KITTI (Behley et al., 2019), Ford (Pandey et al., 2011), and nu Scenes (Caesar et al., 2020). |
| Dataset Splits | Yes | Ford is also collected using Velodyne HDL-64E. The common test condition (CTC) defined by MPEG (WG 07 MPEG 3D Graphics Coding and Haptics Coding, 2024b) utilizes three Ford sequences, each having 1,500 frames at 1mm precision with 8-bit reflectance. The first sequence is for training, and the remaining two are for testing. [...] For training, we extract the first 100 frames from the first 12 scenes in the first five subsets, resulting in 6,000 frames. For testing, we select the first 90 frames from the first scene of each of the last five subsets, yielding 450 frames numbered as sequences #01 to #05. |
| Hardware Specification | Yes | For fair comparisons, all experiments are conducted on the same platform, equipped with an NVIDIA RTX 4090 GPU, an Intel Core i9-13900K CPU, and 64GB of memory. |
| Software Dependencies | Yes | We implement Ser Li C using Python 3.10 and Py Torch 2.5. |
| Experiment Setup | Yes | The model is trained with Adam W optimizer (Loshchilov & Hutter, 2019), using a learning rate of 2 10 4 and a batch size of 64. We employ a cosine annealing strategy (Loshchilov & Hutter, 2017) to gradually reduce the learning rate to 5 10 5. The model is randomly initialized and trained for 25 epochs for each dataset. |