Self-Supervised Robust Scene Flow Estimation via the Alignment of Probability Density Functions
Authors: Pan He, Patrick Emami, Sanjay Ranka, Anand Rangarajan861-869
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comprehensive experiments show that our method makes noticeable gains over the Chamfer Distance and the Earth Mover s Distance in real-world environments and achieves state-of-the-art performance among selfsupervised learning methods on Flying Things3D and KITTI, even outperforming some supervised methods with ground truth annotations. |
| Researcher Affiliation | Academia | Pan He, Patrick Emami, Sanjay Ranka, Anand Rangarajan Department of Computer and Information Science and Engineering, University of Florida 432 Newell Dr, Gainesville, FL 32611, USA pan.he,pemami@ufl.edu; ranka,anand@cise.ufl.edu |
| Pseudocode | No | No clearly labeled "Pseudocode" or "Algorithm" block, or structured pseudocode for the overall method, was found in the main body of the paper. While the paper mentions providing "a few lines of code" in the appendix, this refers to an implementation detail of a formula, not a pseudocode block for the entire approach. |
| Open Source Code | No | No explicit statement about releasing the source code for the methodology described in this paper, nor any link to a code repository, was found. |
| Open Datasets | Yes | Flying Things3D (FT3D): The dataset (Mayer et al. 2016) is the first large-scale synthetic dataset designed for scene flow estimation... KITTI Scene Flow (KSF): This real-world dataset (Menze, Heipke, and Geiger 2015; Menze and Geiger 2015) is adapted from the KITTI Scene Flow benchmark... Unlabelled KITTI๐Dataset: The KITTI๐dataset is prepared by (Li, Lin, and Xie 2021) where they use raw point clouds in KITTI to produce a training dataset. |
| Dataset Splits | Yes | Following (Liu, Qi, and Guibas 2019; Gu et al. 2019), we reconstructed point clouds and their ground truth scene flows, ending up with a total of 19, 640 training examples and 3, 824 test examples. We selected 3, 928 examples from the training set for a hold-out validation. |
| Hardware Specification | No | No specific hardware details (e.g., GPU/CPU models, memory, or detailed computer specifications) used for running experiments were found. The paper only mentions software frameworks used. |
| Software Dependencies | No | The paper mentions "Pytorch" and "Minkowski Engine" but does not specify their version numbers, which are required for a reproducible description of ancillary software. |
| Experiment Setup | Yes | We randomly sampled 8, 192 points for each point cloud. All models were implemented in Pytorch (Paszke et al. 2019) and Minkowski Engine (Choy, Gwak, and Savarese 2019). We utilized the Adam optimizer (Kingma and Ba 2014) with an initial learning rate of 0.001 and the cosine annealing scheduler. We trained models for 100 epochs and voxelized points at a resolution of 0.1m. We empirically set ๐= 0.00625. We empirically set ๐= 50. The full objective function is a weighted sum of the CS divergence loss and the graph Laplacian loss: ๐ธ(D) = ๐ธ๐๐ (D) + ๐๐ธ๐(D) where ๐is a hyperparameter to balance two terms. |