RadarOcc: Robust 3D Occupancy Prediction with 4D Imaging Radar

Authors: Fangqiang Ding, Xiangyu Wen, Yunzhou Zhu, Yiming Li, Chris Xiaoxuan Lu

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on the K-Radar dataset show Radar Occ s state-of-the-art performance in radar-based 3D occupancy prediction and comparable results to other modalities in normal weathers.
Researcher Affiliation Academia 1University of Edinburgh 2Georgia Institute of Technology 3New York University 4AI Centre, Department of Computer Science, UCL
Pseudocode No The paper describes its method through text and a pipeline diagram (Fig. 1) but does not include any explicit pseudocode or algorithm blocks.
Open Source Code Yes We release our code and model at https: //github.com/Toytiny/Radar Occ.
Open Datasets Yes Our experiments are conducted on the K-Radar dataset [42], which is, to the best of our knowledge, the only autonomous driving dataset providing available 4DRT data.
Dataset Splits Yes We split the annotated 24 sequences into the training, validation and test sets with a ratio of 17:2:5, resulting in 11,333, 1,059 and 2,878 frames, respective.
Hardware Specification Yes All of our experiments are conducted on a Ubuntu server equipped with 2 Nvidia RTX 3090 24GB GPUs, an Intel i9-10980XE CPU @ 3.00GHz and a 64GB RAM.
Software Dependencies No The paper mentions software like the 'spconv library' and 'PyTorch' but does not specify their version numbers, which is necessary for reproducible ancillary software details.
Experiment Setup Yes We train Radar Occ with 10 epochs using Adam optimizer with a learning rate of 3e-4. The batch size is 1 for each GPU. We follow [3] to use loss normalization to balance the weight of the 4 different losses, and cosine annealing [107] with 1/3 warm-up ratio is used at the start of the training.