Open-Vocabulary Object Detection Based on Context-Awareness and Hierarchical Classification

In real-world open-scene scenarios, object detection often suffers from bottlenecks such as uneven distribution of tail classes, failure to detect small objects and corruption due to background noise and that is why traditional closed domain detection algorithms will behave badly when presented with new or unlabeled classes. The paper proposes an object detection architecture that can support open vocabulary conditions, named APOVD-Net (Adaptive Prompt Open Vocabulary Detection Network). To address the issue, the network adopts a hierarchical classifier, which follows a coarse-to-fine, large-to-small framework and uses hierarchical semantics routing to provide finer disambiguation. Additionally, a dual-confidence adaptive evaluation is suggested to get a dynamic trade-off in the spatial localization accuracy and cross-modal semantic agreement. Experiments conducted on the official LVIS v1.0 validation set show that APOVD-Net outperforms in both average precision (mAP) and rare category detection rate (Rare AP), although with a very high inference rate, and can be regarded as a viable intelligent visual perception framework for open environments.