SCONE: Surface Coverage Optimization in Unknown Environments by Volumetric Integration
Authors: Antoine Guedon, Pascal Monasse, Vincent Lepetit
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate our approach on a novel dataset made of large and complex 3D scenes. We first compare the performance of our model to the state of the art on the Shape Net dataset [2], following the protocol introduced in [33]. While our method was designed to handle more general frameworks such as 3D scene reconstruction and continuous cameras poses in the scene, it outperforms the state of the art for dense reconstruction of objects when the camera is constrained to stay on a sphere centered on the object. We then conduct experiments in large 3D environments using a simple planning algorithm that builds a camera trajectory online by iteratively selecting NBVs with SCONE. Since, to the best of our knowledge, we propose the first supervised Deep Learning method for such free 6D motion of the camera, we created a dataset made of several large-scale scenes under the CC License for quantitative evaluation. We made our code and this dataset available for allowing comparison of future methods with SCONE on our project webpage: https://github.com/Anttwo/SCONE. We now provide an ablation study for both modules of our full model: The prediction of occupancy probability, and the computation of coverage gain from predictions about visibility gains. |
| Researcher Affiliation | Academia | LIGM, Ecole des Ponts, Univ Gustave Eiffel, CNRS, France |
| Pseudocode | No | The paper does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | Yes | We made our code and this dataset available for allowing comparison of future methods with SCONE on our project webpage: https://github.com/Anttwo/SCONE. |
| Open Datasets | Yes | We first compare the performance of our model to the state of the art on a subset of the Shape Net dataset [2] introduced in [33] and following the protocol of [33]: We sample 4,000 training meshes from 8 specific categories of objects, 400 validation meshes and 400 test meshes from the same categories, and 400 additional test meshes from 8 categories unseen during training. Since, to the best of our knowledge, we propose the first supervised Deep Learning method for such free 6D motion of the camera, we created a dataset made of several large-scale scenes under the CC License for quantitative evaluation. We made our code and this dataset available for allowing comparison of future methods with SCONE on our project webpage: https://github.com/Anttwo/SCONE. |
| Dataset Splits | Yes | We sample 4,000 training meshes from 8 specific categories of objects, 400 validation meshes and 400 test meshes from the same categories, and 400 additional test meshes from 8 categories unseen during training. |
| Hardware Specification | Yes | We iterate the process either 100 times to build a full path around the object in around 5 minutes on a single Nvidia GTX1080 GPU, and recovered a partial point cloud with up to 100,000 points. This work was granted access to the HPC resources of IDRIS under the allocation 2022-AD011013387 made by GENCI. |
| Software Dependencies | No | The paper does not specify any software dependencies with version numbers (e.g., Python version, specific library versions like PyTorch or TensorFlow). |
| Experiment Setup | Yes | We train the occupancy probability module alone with a Mean Squared Error loss with the ground truth occupancy map. We do not compute ground-truth visibility gains to train the second module since it would make computation more difficult and require further assumptions: We supervise directly on IH(c) by comparing it to the ground-truth surface coverage gain for multiple cameras, with softmax normalization and Kullback-Leibler divergence loss. We iterate the process either 100 times to build a full path around the object in around 5 minutes on a single Nvidia GTX1080 GPU, and recovered a partial point cloud with up to 100,000 points. In our experiments, the number of different poses is around 10,000. More details about the experiments can be found in the appendix. Extensive details about the training of our modules and the choices we made are given in the appendix. |