SOGDet: Semantic-Occupancy Guided Multi-View 3D Object Detection

Authors: Qiu Zhou, Jinming Cao, Hanchao Leng, Yifang Yin, Yu Kun, Roger Zimmermann

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

Reproducibility Variable Result LLM Response
Research Type Experimental To evaluate its effectiveness, we apply this approach to several state-of-the-art baselines and conduct extensive experiments on the exclusive nu Scenes dataset. Our results show that SOGDet consistently enhance the performance of three baseline methods in terms of nu Scenes Detection Score (NDS) and mean Average Precision (m AP). We conducted extensive experiments on the nu Scenes (Caesar et al. 2020) dataset
Researcher Affiliation Collaboration Qiu Zhou1*, Jinming Cao2 *, Hanchao Leng3, Yifang Yin4, Yu Kun3, Roger Zimmermann2 1Independent Researcher 2National University of Singapore 3Xiaomi Car 4Institute for Infocomm Research (I2R), A*STAR, Singapore
Pseudocode Yes Algorithm 1: Binary occupancy label generation
Open Source Code Yes The codes are available at: https://github.com/zhouqiu/SOGDet.
Open Datasets Yes We conducted extensive experiments on the nu Scenes (Caesar et al. 2020) dataset, which is currently the exclusive benchmark for both 3D object detection and occupancy prediction.
Dataset Splits Yes Following the standard practice (Huang et al. 2021; Feng et al. 2022), we used the official splits of this dataset: 700 and 150 scenes respectively for training and validation, and the remaining 150 for testing.
Hardware Specification Yes training on eight 80G A100 GPUs with a mini-batch size of 8, for a total batch size of 64
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 We followed the experimental protocol of AEDet (Feng et al. 2022) and training on eight 80G A100 GPUs with a mini-batch size of 8, for a total batch size of 64, and trained the model for 24 epochs with CBGS (Zhu et al. 2019) using Adam W as the optimizer with a learning rate of 2e-4.