Beyond the Label Itself: Latent Labels Enhance Semi-supervised Point Cloud Panoptic Segmentation

Authors: Yujun Chen, Xin Tan, Zhizhong Zhang, Yanyun Qu, Yuan Xie

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The ablation of the IPSL module demonstrates its robust adaptability, and the experiments evaluated on Semantic KITTI and nu Scenes demonstrate that our model outperforms the state-of-the-art method, Laser Mix.
Researcher Affiliation Academia 1School of Computer Science and Technology, East China Normal University 2Chongqing Institute, East China Normal University 3School of Information Science and Engineering, Xiamen University
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide concrete access to source code for the methodology described.
Open Datasets Yes Semantic KITTI (Behley, Milioto, and Stachniss 2021) contains 10 (1/11) training (validation/testing) sequences and totally 43551 Li DAR scans with a 64-beam Li DAR sensor, as well as binocular camera images of each scan additionally. There are point-wise panoptic annotations with 8 thing class and 12 stuff class labels with instance labels. nu Scenes (Caesar et al. 2020) is a large-scale autopilot dataset with various urban scenes and a 32-beam Li DAR sensor.
Dataset Splits Yes We adopt a sampling method of selecting frames with fixed intervals, which is the same as Kong et al. (2023). In this paper, we have respectively selected 40%, 20%, 10%, and 1% point cloud frames with point-wise labels for training. Table 1: Segmentation results compared with other semi-supervised methods on validation set of nu Scenes and Semantic KITTI.
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper mentions software like MMdetection, Swift Net, Cylinder3D, Panoptic Polar Net, SAM, and Grounded-Segment-Anything, but does not specify their version numbers.
Experiment Setup Yes Implementation Default settings unless ablation. In IPSL Module, the radius of Gaussian heatmaps R = 5, percentage of scale to the center Pcenter = 1/4. For Cylinder Mix, the region size [Rx, Ry, Rz] = [4, 4, 2], the probability pclymix = 25%. For the weight of each term in loss in Lseg, µhm = 100, µos = 10, µfm = 1.