Towards Learning Group-Equivariant Features for Domain Adaptive 3D Detection
Authors: Sangyun Shin, Yuhang He, Madhu Vankadari, Ta-Ying Cheng, Qian Xie, Andrew Markham, Niki Trigoni
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive adaptation experiments on KITTI, Waymo, and Nu Scenes datasets show the effectiveness of our approach for bridging inter-domain gaps. |
| Researcher Affiliation | Academia | Department of Computer Science, University of Oxford, United Kingdom School of Computer Science, University of Leeds, United Kingdom |
| Pseudocode | Yes | Algorithm 1 Training Pipeline |
| Open Source Code | Yes | https://github.com/yunshin/Group Exp-DA.git |
| Open Datasets | Yes | We evaluate our methods against various baselines across three different datasets, such as KITTI [6], Nu Scenes [1], and Waymo [33]. |
| Dataset Splits | Yes | KITTI contains 7481 frames of point clouds for training and validation, and all the data is collected with 64-beam Velodyne Li DAR. Nu Scenes dataset contains 28130 training and 6019 validation point clouds collected with a 32-beam roof Li DAR. Waymo dataset contains 122000 training and 30407 validation frames of point clouds collected with five Li DAR sensors, i.e., one 64-beam Li DAR and four 200-beam Li DAR. |
| Hardware Specification | Yes | Following [17], we first train each detector for 50 epochs with batch-size 8 as a pretraining step using a single NVIDIA A10 GPU. |
| Software Dependencies | No | The paper mentions 'Open PCDet [35]' as the implementation basis but does not provide specific version numbers for any software dependencies. |
| Experiment Setup | Yes | For a fair comparison with existing domain adaptive 3D detection methods, we build our model on two base detectors, Second Io U [42] and Point Pillars [12], following [44, 43, 17, 8], that are widely used and applicable to most recent detectors with the implementation based on Open PCDet [35] and parameters from ST3D [44]. Following [17], we first train each detector for 50 epochs with batch-size 8 as a pretraining step using a single NVIDIA A10 GPU. In the self-training stage, we train 30 more epochs for the tuning to adapt to the target domain. The learning rate is set to 1 15 4 using Adam optimizer with Cosine annealing [47] for scheduling the learning rate. ... The feature dimensions dbev and dobj for Fbev and µ are both set to 512 for c Attn. α for updating the group parameters is empirically set to 0.8. λ1, λ2, and λ3 are set to 0.5, 0.5, and 1.0, respectively. |