Sparse2Dense: Learning to Densify 3D Features for 3D Object Detection
Authors: Tianyu Wang, Xiaowei Hu, Zhengzhe LIU, Chi-Wing Fu
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our method on the large-scale Waymo Open Dataset and the Waymo Domain Adaptation Dataset, showing its high performance and efficiency over the state of the arts. The code is available at https://github.com/stevewongv/Sparse2Dense. We evaluate our approach on the large-scale benchmark datasets, Waymo [18] open and Waymo domain adaptation, demonstrating its superior performance over the state of the arts. |
| Researcher Affiliation | Academia | Tianyu Wang1,2,3, Xiaowei Hu3, , Zhengzhe Liu1, Chi-Wing Fu1,2 1 The Chinese University of Hong Kong 2 The Shun Hing Institute of Advanced Engineering 3 Shanghai AI Laboratory {wangty,zzliu,cwfu}@cse.cuhk.edu.hk, huxiaowei@pjlab.org.cn |
| Pseudocode | No | The paper describes methods in text and uses figures but does not contain a structured pseudocode or algorithm block. |
| Open Source Code | Yes | The code is available at https://github.com/stevewongv/Sparse2Dense. |
| Open Datasets | Yes | We employ the Waymo Open Dataset and the Waymo Domain Adaptation Dataset [18], which are under the Waymo Dataset License Agreement, to evaluate our framework. Waymo Open Dataset is the largest and most informative 3D object detection dataset, which includes 360 Li DAR point cloud and annotated 3D bounding boxes. The training set contains 798 sequences with around 158K Li DAR frames and the validation set includes 202 sequences with around 40k Li DAR frames. |
| Dataset Splits | Yes | The training set contains 798 sequences with around 158K Li DAR frames and the validation set includes 202 sequences with around 40k Li DAR frames. Following [4, 14, 30], we adopt 20% subset of the Waymo Open Dataset to train our models. |
| Hardware Specification | Yes | We train the DDet on four Nvidia RTX 3090 GPUs with a batch size of four per GPU for 30 epochs. |
| Software Dependencies | No | The paper does not provide specific version numbers for software dependencies such as libraries or frameworks (e.g., PyTorch, TensorFlow, CUDA). |
| Experiment Setup | Yes | In the first training stage, following [1], we train DDet from scratch using Adam with a learning rate of 0.003 and a one-cycle learning rate policy with a dividing factor of 0.1 and a percentage of the cycle of 0.3. We set the detect range as [ 75.2m, 75.2m] for the X, Y axes and set [ 2m, 4m] for the Z axis, and the size of each voxel grid as (0.1m, 0.1m, 0.15m). We apply global rotation around the Z-axis, random flipping, global scaling, and global translating as the data augmentation. We train the DDet on four Nvidia RTX 3090 GPUs with a batch size of four per GPU for 30 epochs. |