OLIVE: Object Level In-Context Visual Embeddings
This addresses the limitation of existing VLMs in handling domain-specific tasks and unseen visual concepts, offering a more efficient and robust approach for object-level reasoning.
The paper tackles the problem of fine-grained object-level understanding and grounding in vision-language models by proposing a method to prompt large language models with in-context visual object vectors, achieving competitive performance in referring object classification and captioning with zero-shot generalization.
Recent generalist vision-language models (VLMs) have demonstrated impressive reasoning capabilities across diverse multimodal tasks. However, these models still struggle with fine-grained object-level understanding and grounding. In terms of modeling, existing VLMs implicitly align text tokens with image patch tokens, which is ineffective for embedding alignment at the same granularity and inevitably introduces noisy spurious background features. Additionally, these models struggle when generalizing to unseen visual concepts and may not be reliable for domain-specific tasks without further fine-tuning. To address these limitations, we propose a novel method to prompt large language models with in-context visual object vectors, thereby enabling controllable object-level reasoning. This eliminates the necessity of fusing a lengthy array of image patch features and significantly speeds up training. Furthermore, we propose region-level retrieval using our object representations, facilitating rapid adaptation to new objects without additional training. Our experiments reveal that our method achieves competitive referring object classification and captioning performance, while also offering zero-shot generalization and robustness to visually challenging contexts.