SL-Ground
A sentence-level grounding model for open-vocabulary object detection that encodes category descriptions as complete sentence-level representations and directly aligns and scores them against object queries, replacing traditional token-level phrase grounding frameworks and improving detection performance on fine-grained compositional categories.
Timeline (last 90 days)
SL-Ground proposes encoding each category description as a complete sentence-level representation, letting object queries score directly against the holistic representation rather than detouring through tokens.
SL-Ground avoids the need for prompt chunking through whole-sentence encoding, simplifying the deployment process and reducing engineering complexity on long-tail categories.
The authors built a multi-granularity Stanford Cars benchmark covering three levels: vehicle type, brand, and specific model.
At the model granularity, SL-Ground raises AP from 14.5 for the token-level variant to 70.6 (under matched training candidate sampling conditions).
When trained with the full vocabulary, SL-Ground further reaches an AP of 86.0 at the model granularity.
On the FG-OVD benchmark, SL-Ground achieves the best average AP among compared methods.
Current evaluation is concentrated on the automotive domain and the FG-OVD benchmark, and whether sentence-level encoding is equally robust under more open and noisier description distributions remains to be verified.
All Facts (7)
SL-Ground proposes encoding each category description as a complete sentence-level representation, letting object queries score directly against the holistic representation rather than detouring through tokens.
50%UnverifiedSL-Ground avoids the need for prompt chunking through whole-sentence encoding, simplifying the deployment process and reducing engineering complexity on long-tail categories.
50%UnverifiedThe authors built a multi-granularity Stanford Cars benchmark covering three levels: vehicle type, brand, and specific model.
50%UnverifiedAt the model granularity, SL-Ground raises AP from 14.5 for the token-level variant to 70.6 (under matched training candidate sampling conditions).
50%UnverifiedWhen trained with the full vocabulary, SL-Ground further reaches an AP of 86.0 at the model granularity.
50%UnverifiedOn the FG-OVD benchmark, SL-Ground achieves the best average AP among compared methods.
50%UnverifiedCurrent evaluation is concentrated on the automotive domain and the FG-OVD benchmark, and whether sentence-level encoding is equally robust under more open and noisier description distributions remains to be verified.
50%