Metoda za detekcijo objektov s tekstovnim poizvedovanjem

The increasing diversity and scale of object detection datasets have highlighted the limitations of closed-set detectors with fixed vocabularies. Open-set object detection addresses this by enabling the detection of arbitrary classes via text prompts. Grounding DINO is a prominent zero-shot detector, but its training code is not fully open-source, and its implementation is outdated. In this work, we reimplement Grounding DINO, achieving ~20% speedup, and extend it for text-based object counting, though these modifications do not consistently improve FSCD-147 performance. To enable training from scratch, we optimize the model further, achieving an additional ~30% speedup, and develop an edge-oriented variant inspired by a closed-source model. Training on 1.3 million images from scratch, we evaluate the model on COCO and LVIS datasets, where it performs comparably to other open-source models, but remains below the closed-source baseline, likely due to a much smaller training set.