OWL-ViT

Open-vocabulary detection is the point of OWL-ViT: instead of a fixed list of classes, you describe what to find in plain text, or supply an example crop of the object, and the model returns boxes with confidence scores. Google Research built it from CLIP-style image and text encoders plus lightweight box and class heads, so detection inherits the broad vocabulary learned from web image-text pairs. A second version, OWLv2, added self-training on web data and scores clearly higher on rare LVIS categories, with checkpoints that drop into the same code. The reference implementation sits in Google's Scenic JAX/Flax repository with ViT-B/32, B/16 and L/14 checkpoints, Colab notebooks for inference and TensorFlow export, plus training and evaluation scripts. Most developers skip JAX and load the OWL-ViT or OWLv2 weights through Hugging Face Transformers in PyTorch. Base models run on a modest GPU; the L/14 checkpoints, around 1.7 GB on disk, want more memory. Code and weights are Apache-2.0, so commercial products may use them. Handy for zero-shot labeling and image search when training a custom detector is not worth it.