Learning Object Bounding Boxes for 3D Instance Segmentation on Point Clouds
Authors: Bo Yang, Jianan Wang, Ronald Clark, Qingyong Hu, Sen Wang, Andrew Markham, Niki Trigoni
NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that our approach surpasses existing work on both Scan Net and S3DIS datasets while being approximately 10 more computationally efficient. Comprehensive ablation studies demonstrate the effectiveness of our design. |
| Researcher Affiliation | Collaboration | 1University of Oxford 2Deep Mind 3Imperial College London 4Heriot-Watt University |
| Pseudocode | Yes | Algorithm 1 An algorithm to calculate point-in-pred-boxprobability. |
| Open Source Code | Yes | Our code and data are available at https://github.com/Yang7879/3D-Bo Net. |
| Open Datasets | Yes | We first evaluate our approach on Scan Net(v2) 3D semantic instance segmentation benchmark [7]. Similar to SGPN [51], we divide the raw input point clouds into 1m 1m blocks for training, while using all points for testing followed by the Block Merging algorithm [51] to assemble blocks into complete 3D scenes. We further evaluate the semantic instance segmentation of our framework on S3DIS [1], which consists of 3D complete scans from 271 rooms belonging to 6 large areas. |
| Dataset Splits | Yes | We divide the raw input point clouds into 1m 1m blocks for training, while using all points for testing followed by the Block Merging algorithm [51] to assemble blocks into complete 3D scenes. In our experiments, H is set as 24 and we follow the 6-fold evaluation [1; 52]. |
| Hardware Specification | Yes | The whole network is trained on a Titan X GPU from scratch. |
| Software Dependencies | No | The paper mentions using "Adam solver" and "focal loss" and relies on PointNet++ as a backbone but does not specify version numbers for any software libraries or frameworks (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | We use Adam solver [19] with its default hyper-parameters for optimization. Initial learning rate is set to 5e 4 and then divided by 2 every 20 epochs. The whole network is trained on a Titan X GPU from scratch. We use the same settings for all experiments, which guarantees the reproducibility of our framework. We use focal loss [30] with default hyper-parameters instead of the standard cross-entropy loss to optimize this branch. H is set as 24. We use θ1 = 100, θ2 = 20 in all our implementation. |