GFNet: Geometric Flow Network for 3D Point Cloud Semantic Segmentation
Authors: Haibo Qiu, Baosheng Yu, Dacheng Tao
TMLR 2022 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform extensive experiments on two widely used benchmark datasets, Semantic KITTI and nu Scenes, to demonstrate the effectiveness of our GFNet for project-based point cloud semantic segmentation. Concretely, GFNet not only significantly boosts the performance of each individual view but also achieves state-of-the-art results over all existing projectionbased models. |
| Researcher Affiliation | Academia | Haibo Qiu EMAIL School of Computer Science, The University of Sydney, Australia. Baosheng Yu EMAIL School of Computer Science, The University of Sydney, Australia. Dacheng Tao EMAIL School of Computer Science, The University of Sydney, Australia |
| Pseudocode | Yes | Algorithm 1 Geometric Flow Module (BEV RV) |
| Open Source Code | Yes | Code is available at https://github.com/haibo-qiu/GFNet. |
| Open Datasets | Yes | Semantic KITTI (Behley et al., 2019), derived from the KITTI Vision Benchmark (Geiger et al., 2012), provides dense point-wise annotations for semantic segmentation task... nu Scenes (Caesar et al., 2020) is a large-scale autonomous driving dataset, containing 1000 driving scenes of 20 second length in Boston and Singapore. |
| Dataset Splits | Yes | Semantic KITTI... these 22 sequences are divided into 3 sets, i.e., training set (00 to 10 except 08 with 19130 scans), validation set (08 with 4071 scans) and testing set (11 to 21 with 20351 scans). nu Scenes... officially divided into training (850 scenes) and validation set (150 scenes). |
| Hardware Specification | Yes | We train the proposed GFNet for 150 epochs using the batch size 16 on four NVIDIA A100-SXM4-40GB GPUs with AMD EPYC 7742 64-Core Processor... train the model for total 400 epoch with batch size 56 using 8 NVIDIA A100-SXM4-40GB GPUs under AMD EPYC 7742 64-Core Processor. |
| Software Dependencies | No | The paper mentions software like SGD optimizer, cosine learning rate schedule, and pretrained weights from ImageNet, but does not provide specific version numbers for any software libraries or frameworks used (e.g., Python, PyTorch, TensorFlow, CUDA versions). |
| Experiment Setup | Yes | For Semantic KITTI, we use two branches to learn representations from RV/BEV in an end-to-end trainable way, where each branch follows an encoder-decoder architecture with a Res Net-34 (He et al., 2016) as the backbone. The ASPP module (Chen et al., 2017b) is also used between the encoder and the decoder... We employ a SGD optimizer with momentum 0.9 and the weight decay 1e 4. We use the cosine learning rate schedule (Loshchilov & Hutter, 2016) with warmup at the first epoch to 0.1... By default, we use λ = [2.0, 2.0, 2.0, 1.0, 1.0] as the loss weight for Eq.11. We train the proposed GFNet for 150 epochs using the batch size 16... |