Towards Efficient 3D Object Detection with Knowledge Distillation
Authors: Jihan Yang, Shaoshuai Shi, Runyu Ding, Zhe Wang, Xiaojuan Qi
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, we conduct extensive experiments on the Waymo and KITTI dataset. Our best performing model achieves 65.75% LEVEL 2 m APH, surpassing its teacher model and requiring only 44% of teacher flops on Waymo. |
| Researcher Affiliation | Collaboration | Jihan Yang1 Shaoshuai Shi2 Runyu Ding1 Zhe Wang3 Xiaojuan Qi1 1The University of Hong Kong 2Max Planck Institute for Informatics 3Sense Time Research |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code is available at https://github.com/CVMI-Lab/Sparse KD. |
| Open Datasets | Yes | For dataset, we perform all experiments on the largest annotated 3D LiDAR perception dataset Waymo Open Dataset (WOD) [37] with 20% training samples for fast verification. In addition, our method can generalize well to other settings such as KITTI dataset with SECOND as well as advance compression methods in Sec. 5.5, other detectors, and even 3D semantic segmentation in Suppl.. |
| Dataset Splits | Yes | For dataset, we perform all experiments on the largest annotated 3D LiDAR perception dataset Waymo Open Dataset (WOD) [37] with 20% training samples for fast verification. |
| Hardware Specification | Yes | Most of experiments are trained with 8 NVIDIA 1080Ti. Few experiments are trained with 8 NVIDIA V100 or NVIDIA A100. Full set results on Waymo are trained with 16 NVIDIA 1080ti. |
| Software Dependencies | No | The paper mentions following the training scheme of popular 3D detection codebase Open PCDet [41], but does not provide specific version numbers for software dependencies such as Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | For model training, we follow the training scheme of popular 3D detection codebase Open PCDet [41] to ensure fair comparisons and standardization. Implementation details and the value of hyper-parameters are described in the Suppl.. |