OWL-ViT image-conditioned input: Is that possibale to input multiple images?

[liu-bohan]

I want to detect multiple image-conditioned queries on a single image at one time.

I use the code from OWL_ViT_minimal_example.ipynb for Image-conditioned detection.

target_class_predictions = class_predictor(
    image_features=feature_map.reshape(b, h * w, d),
    query_embeddings=query_embedding[None, None, ...],  # [batch, queries, d]
)

It looks like I can input multiple image queries with only one query_embedding. But, after I get predicted logits and bboxes, I can't figure out corresponding labels and bboxes.