Core Concepts
The author introduces AmodalSynthDrive, a synthetic dataset for amodal perception in autonomous driving, addressing the challenges of occlusion reasoning and depth estimation.
Abstract
AmodalSynthDrive is a comprehensive dataset providing multi-view camera images, 3D bounding boxes, LiDAR data, and odometry for various tasks in amodal perception. It aims to advance research in autonomous driving by offering benchmarks and novel tasks like amodal depth estimation.
The dataset facilitates the development of standalone and integrated approaches for amodal scene understanding tasks. It includes annotations for traditional modal perception tasks as well.
Key challenges include predicting occluded regions accurately and modeling relative occlusion order for precise depth estimations. The dataset's complexity is highlighted by its diverse weather conditions and detailed occluded region annotations.
Various baselines are evaluated on the dataset to demonstrate challenges and improvements in tasks like amodal panoptic segmentation, instance segmentation, and semantic segmentation.
Transfer learning results show that pre-training on AmodalSynthDrive enhances performance on real-world datasets for amodal instance segmentation and panoptic segmentation.
Stats
The dataset provides over 1M object annotations in diverse traffic, weather, and lighting conditions.
The training set encompasses 105 video sequences with 42,000 images while the test set holds 30 video sequences representing 12,000 images.
The dataset consists of 18 distinct semantic classes with instance annotations provided for 7 classes.
AmodalSynthDrive supports multiple tasks including amodal semantic segmentation, instance segmentation, panoptic tracking, motion segmentation, and more.