Dissecting, disclosing as well as diminishing uncertainty encoded in processed data can help to protect stakeholders from harm during the deployment of machine learning systems. Enforcing annotator agreement, ensemble methods, post-hoc calibration or rejection of adversarial input are commonly proposed to increase safety. ISeVis is introduced, e.g., to investigate the relationship between noisy labels and model voting. Here, classification-type obstacle avoidance is treated as the exemplary computer vision task to be accomplished. Besides path planning or localization and mapping, avoiding collisions constitutes moreover a prerequisite in the broader realm of navigation. Hence, obstacle avoidance becomes increasingly relevant in critical and high-risk scenarios, such as automated transportation. ISeVis encompasses one hour of camera footage from a ground-based vehicle where two distinct parking settings are encountered. Different lighting conditions, varying camera angles and randomly appearing objects are considered. Beyond obstacle avoidance and because of its visual similarity, ISeVis may complement other widely utilized datasets, such as NYU (depth estimation) or COCO (object detection/segmentation).
ISeVis: Indoor Scene Visual Obstacle Avoidance
Research Data
Electronic Resource
Unknown
Indoor navigation and obstacle avoidance for unmanned aerial vehicles
European Patent Office | 2025
|Chassis obstacle avoidance system and obstacle avoidance method
European Patent Office | 2023
|European Patent Office | 2020
|European Patent Office | 2021
|