Hindsight is 20/20: Leveraging Past Traversals to Aid 3D Perception

Authors: Yurong You, Katie Z Luo, Xiangyu Chen, Junan Chen, Wei-Lun Chao, Wen Sun, Bharath Hariharan, Mark Campbell, Kilian Q Weinberger

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate our approach on two large-scale, real-world self-driving datasets, Lyft Level 5 Perception (Kesten et al., 2019) and the nu Scenes Dataset (Caesar et al., 2020), with multiple representative modern object detection models (Lang et al., 2019; Yan et al., 2018; Zhou & Tuzel, 2018; Shi et al., 2019; 2020a) under various settings and show consistent performance gains. Our code is available at https://github.com/Yurong You/Hindsight. 4 EXPERIMENTS
Researcher Affiliation Academia 1Cornell University, Ithaca, NY 2The Ohio State University, Columbus, OH
Pseudocode No The paper describes the pipeline and components but does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes Our code is available at https://github.com/Yurong You/Hindsight.
Open Datasets Yes We validate our approach, HINDSIGHT, on the Lyft Level 5 Perception Dataset (Kesten et al., 2019) (Lyft) and the nu Scenes Dataset (Caesar et al., 2020).
Dataset Splits No The paper states training/test splits (e.g., "12,407/2,274 training/test samples"), but it does not explicitly specify a separate validation dataset split.
Hardware Specification Yes All models are trained with 4 NVIDIA 3090 GPUs.
Software Dependencies No The paper mentions using "Minkowski Engine" and "Open PCDet" but does not provide specific version numbers for these software dependencies, which are necessary for full reproducibility.
Experiment Setup Yes We set [ Hs, He], the range of scans that are combined to produce dense point clouds, to be [0, 20] m... We use default quantization size δ = 0.3 m for all the experiments... Due to GPU memory limit in training, we use Tmax = 5 available past traversals... The filter kernel size K is 5. The dimensionality of the history features dhistory is 64... We use most of the default hyper-parameters tuned for KITTI, with the exception that we enlarge the perception range from 70m to 90m... and we reduce the number of training epochs by 1/4 on Lyft.