Neural Lighting Simulation for Urban Scenes

Authors: Ava Pun, Gary Sun, Jingkang Wang, Yun Chen, Ze Yang, Sivabalan Manivasagam, Wei-Chiu Ma, Raquel Urtasun

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments show that Light Sim generates more realistic relighting results than prior work. Importantly, training perception models on data generated by Light Sim can significantly improve their performance.
Researcher Affiliation Collaboration Waabi1 University of Toronto2 University of Waterloo3 MIT4
Pseudocode No The paper describes methods in text and uses diagrams, but does not include structured pseudocode or algorithm blocks.
Open Source Code No Our project page is available at https://waabi.ai/lightsim/. (This is a project page, not a direct link to the source code for the proposed method.)
Open Datasets Yes We evaluate our method primarily on the public real-world driving dataset Panda Set [87] ... To showcase generalizability, we also demonstrate our approach on ten dynamic scenes from the nu Scenes [11] dataset.
Dataset Splits Yes Specifically, we train on 68 snippets collected in the city and evaluate on 35 snippets in a suburban area, since these two collections are independent and exposed to different lighting conditions.
Hardware Specification Yes In this project, we ran the experiments primarily on NVIDIA Tesla T4s provided by Amazon Web Services (AWS). For prototype development and small-scale experiments, we used local workstations with RTX A5000s.
Software Dependencies No We use the official repository for training and evaluating our model on Panda Set. ... Models were trained for five epochs using the Adam W optimizer [48], coupled with the cosine learning rate schedule.8 (No specific version numbers for software dependencies are provided.)
Experiment Setup Yes We adopt the BEVFormer-small architecture7 with a batch size of two per GPU. Models were trained for five epochs using the Adam W optimizer [48], coupled with the cosine learning rate schedule.