Active 3D Shape Reconstruction from Vision and Touch

Authors: Edward Smith, David Meger, Luis Pineda, Roberto Calandra, Jitendra Malik, Adriana Romero Soriano, Michal Drozdzal

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments show the benefits of such solutions in the task of 3D shape understanding where our models consistently outperform natural baselines. We provide our framework as a tool to foster future research in this direction.
Researcher Affiliation Collaboration Edward J. Smith1,2 David Meger2 Luis Pineda1 Roberto Calandra1 Jitendra Malik1,3 Adriana Romero-Soriano1,2, Michal Drozdzal1, 1 Facebook AI Research 2 Mc Gill University 3 University of California, Berkeley
Pseudocode No The paper describes the steps of the simulator and the reconstruction pipeline conceptually (e.g., Figure 1, Figure 3, detailed paragraphs) but does not include any formal pseudocode or algorithm blocks.
Open Source Code Yes Our framework, training and evaluation setup, and trained models are publicly available on a Git Hub repository to ensure and encourage reproducible experimental comparison 3. 3https://github.com/facebookresearch/Active-3D-Vision-and-Touch
Open Datasets Yes The dataset used is made up of 40,000 objects sampled from the ABC dataset [34, 56], a CAD model dataset of approximately one million objects.
Dataset Splits Yes This set of objects was split into 5 sets; 3 training sets 5 of size 7,700 object each, a validation set comprised 2,000 objects, and a test set of size 1,000.
Hardware Specification Yes All steps in this procedure are performed in parallel or using GPU accelerated computing, and as a result across the 50 grasping options of 100 randomly chosen objects, simulated grasps and touch signals are produced in 0.0317 seconds each on a Tesla V100 GPU with 16 CPU cores.
Software Dependencies No In our simulator, all steps are performed in python across the robotics simulator Py Bullet [15], the rendering tool Pyrender [39], and Py Torch [48]. The paper lists software but does not specify version numbers for reproducibility.
Experiment Setup No The paper mentions aspects of the experimental setup such as number of grasps (5 grasps) and performing hyper-parameter search, but it does not provide specific numerical values for hyperparameters like learning rates, batch sizes, or optimizer settings in the main text.