Towards Optimal Strategies for Training Self-Driving Perception Models in Simulation
Authors: David Acuna, Jonah Philion, Sanja Fidler
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In Section 6, we experimentally validate the efficacy of our method and show the same approach can be simply applied to different sensors and data modalities (oftentimes implying different architectures). Although we focus on the scenario where labeled data is not available in the target domain, we show our framework also provides gains when labels in the target domain are available. Finally, we show that variations such as number of vehicle assets, map design and camera post-processing effects can be compensated for with our method, thus showing what variations are less important to include in the simulator. |
| Researcher Affiliation | Collaboration | David Acuna , Jonah Philion , Sanja Fidler NVIDIA, University of Toronto, Vector Institute {dacunamarrer, jphilion, sfidler}@nvidia.com |
| Pseudocode | No | The paper references "pseudocode in appendix" in Figure 2, but no pseudocode block or algorithm section is included in the provided text of the paper itself. |
| Open Source Code | No | The datasets and scripts for generating them will be publicly released. |
| Open Datasets | Yes | We showcase our approach on the bird s-eye-view vehicle segmentation task with multi-sensor data (cameras, lidar) using an open-source simulator (CARLA), and evaluate the entire framework on a real-world dataset (nu Scenes). |
| Dataset Splits | Yes | We split each dataset into a training set of 28k timesteps (which matches the number of timesteps in the nu Scenes training set) and a validation set of 4k timesteps. |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., GPU models, CPU types, memory) used to run the experiments. |
| Software Dependencies | No | The paper mentions "CARLA version 0.9.11" as the simulator used. While it mentions other software/models like "Lift Splat method [37]" and "Point Pillars [29]", it does not provide specific version numbers for these or other key software dependencies (e.g., deep learning frameworks like PyTorch or TensorFlow, and their versions) used in their implementation. |
| Experiment Setup | No | The paper mentions specific details like "We let τ = 0.9 in all our experiments." and that they use the "same backbone [49] and training scheme [28, 9] as in [37]". However, it does not provide a comprehensive set of specific hyperparameter values such as learning rates, batch sizes, number of epochs, or detailed optimizer settings necessary to fully reproduce the experimental setup. |