8 February 2024
Computer Vision and Pattern Recognition
Proposes HIL method to refine grasp dataset by manually fixing errors
Humans categorize low-confidence predictions as missing/wrong labels
Missing labels get added, wrong labels get removed from dataset
Iterative process enhances dataset quality and model accuracy
Various networks gain 5-8% better grasp detection without architecture changes
Enhancing Robotic Grasping Dataset Quality with Human Oversight
This paper proposes a Human-in-the-Loop (HIL) approach to refine an existing robotic grasping dataset by having humans evaluate and correct errors in automated grasp predictions. Low-confidence predictions are manually categorized and corrected, iteratively improving dataset quality. After multiple rounds, accuracy for various networks increased 5-8% without changing model architectures, demonstrating the impact of higher-quality training data.
No comments yet, be the first to start the conversation...
Sign up to comment on this paper