Benchmark evaluation demonstrates superior novel object detection and inference speed in an assembly scenario, highlighting effective adaptation without model fine-tuning.
Key Points
Zero-shot novel item identification outperforms the DE-ViT baseline in detection capability and processing speed during industrial human-robot collaboration.
Benchmark testing pairs a class-agnostic detector with discriminative embeddings from object re-identification to identify unseen categories from few shots.
Predicting regions of interest through learned embeddings narrows the detector search space, enabling rapid inference without runtime fine-tuning.