Learning Superpoint Graph Cut for 3D Instance Segmentation

Authors: Le Hui, Linghua Tang, Yaqi Shen, Jin Xie, Jian Yang

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on two challenging datasets, Scan Net v2 and S3DIS, show that our method achieves new state-of-the-art performance on 3D instance segmentation. Code is available at https://github.com/fpthink/Graph Cut.
Researcher Affiliation Academia Le Hui , Linghua Tang , Yaqi Shen, Jin Xie , Jian Yang PCA Lab, Nanjing University of Science and Technology, China {le.hui, tanglinghua, syq, csjxie, csjyang}@njust.edu.cn
Pseudocode Yes Algorithm 1 Proposal Generation Algorithm Input: node semantic scores S = {s1, . . . , s|V | | si RN for i = 1, . . . , |V |}, N is the number of classes; semantic threshold θ; edge scores A = {au,v} R|E| 1, au,v indicates the score of edge which connects nodes u and v; Output: proposals I = {I1, . . . , Im}, m is the number of proposals.
Open Source Code Yes Code is available at https://github.com/fpthink/Graph Cut.
Open Datasets Yes Datasets. We conduct experiments on two benchmark datasets, Scan Net v2 [7] and S3DIS [1].
Dataset Splits Yes The Scan Net v2 dataset contains 1,613 3D scenes, which are split into 1,201 training, 312 validation, and 100 test scenes, respectively.
Hardware Specification Yes Our model is trained on a single TITAN RTX GPU.
Software Dependencies No The paper mentions using the Adam optimizer and submanifold sparse convolution, but it does not specify any software names with version numbers (e.g., PyTorch, TensorFlow, CUDA versions).
Experiment Setup Yes Our model is trained on a single TITAN RTX GPU. We use the Adam optimizer with a base learning rate of 0.001 for the network training, which is scheduled by a cosine annealing. The voxel size is set to 0.02m. ... At training time, we limit the maximum number of points in a scene to 250k and crop the excess randomly. Due to the high point density of S3DIS, we randomly downsample its 1/4 points before cropping.