Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Interpreting Object-level Foundation Models via Visual Precision Search

Ruoyu Chen, Siyuan Liang, Jingzhi Li, Shiming Liu, Maosen Li, Zhen Huang, Hua Zhang, Xiaochun Cao

TL;DR

This work tackles the interpretability of object-level foundation models by introducing Visual Precision Search (VPS), a gradient-free, region-sparse attribution method grounded in submodular optimization. By partitioning input images into sparse superpixel groups and scoring regions via a combined Clue and Collaboration metric, VPS yields precise, instance-level saliency maps with theoretical guarantees and practical efficiency. Across Grounding DINO and Florence-2, VPS achieves state-of-the-art faithfulness and localization on COCO, RefCOCO, and LVIS, and enables systematic analysis of grounding and detection failures. The approach demonstrates robust interpretability gains and broad applicability to multimodal detectors, with insights into how input-level factors drive errors and how explanations can guide model improvement.

Abstract

Advances in multimodal pre-training have propelled object-level foundation models, such as Grounding DINO and Florence-2, in tasks like visual grounding and object detection. However, interpreting these models' decisions has grown increasingly challenging. Existing interpretable attribution methods for object-level task interpretation have notable limitations: (1) gradient-based methods lack precise localization due to visual-textual fusion in foundation models, and (2) perturbation-based methods produce noisy saliency maps, limiting fine-grained interpretability. To address these, we propose a Visual Precision Search method that generates accurate attribution maps with fewer regions. Our method bypasses internal model parameters to overcome attribution issues from multimodal fusion, dividing inputs into sparse sub-regions and using consistency and collaboration scores to accurately identify critical decision-making regions. We also conducted a theoretical analysis of the boundary guarantees and scope of applicability of our method. Experiments on RefCOCO, MS COCO, and LVIS show our approach enhances object-level task interpretability over SOTA for Grounding DINO and Florence-2 across various evaluation metrics, with faithfulness gains of 23.7%, 31.6%, and 20.1% on MS COCO, LVIS, and RefCOCO for Grounding DINO, and 102.9% and 66.9% on MS COCO and RefCOCO for Florence-2. Additionally, our method can interpret failures in visual grounding and object detection tasks, surpassing existing methods across multiple evaluation metrics. The code will be released at https://github.com/RuoyuChen10/VPS.

Interpreting Object-level Foundation Models via Visual Precision Search

TL;DR

This work tackles the interpretability of object-level foundation models by introducing Visual Precision Search (VPS), a gradient-free, region-sparse attribution method grounded in submodular optimization. By partitioning input images into sparse superpixel groups and scoring regions via a combined Clue and Collaboration metric, VPS yields precise, instance-level saliency maps with theoretical guarantees and practical efficiency. Across Grounding DINO and Florence-2, VPS achieves state-of-the-art faithfulness and localization on COCO, RefCOCO, and LVIS, and enables systematic analysis of grounding and detection failures. The approach demonstrates robust interpretability gains and broad applicability to multimodal detectors, with insights into how input-level factors drive errors and how explanations can guide model improvement.

Abstract

Advances in multimodal pre-training have propelled object-level foundation models, such as Grounding DINO and Florence-2, in tasks like visual grounding and object detection. However, interpreting these models' decisions has grown increasingly challenging. Existing interpretable attribution methods for object-level task interpretation have notable limitations: (1) gradient-based methods lack precise localization due to visual-textual fusion in foundation models, and (2) perturbation-based methods produce noisy saliency maps, limiting fine-grained interpretability. To address these, we propose a Visual Precision Search method that generates accurate attribution maps with fewer regions. Our method bypasses internal model parameters to overcome attribution issues from multimodal fusion, dividing inputs into sparse sub-regions and using consistency and collaboration scores to accurately identify critical decision-making regions. We also conducted a theoretical analysis of the boundary guarantees and scope of applicability of our method. Experiments on RefCOCO, MS COCO, and LVIS show our approach enhances object-level task interpretability over SOTA for Grounding DINO and Florence-2 across various evaluation metrics, with faithfulness gains of 23.7%, 31.6%, and 20.1% on MS COCO, LVIS, and RefCOCO for Grounding DINO, and 102.9% and 66.9% on MS COCO and RefCOCO for Florence-2. Additionally, our method can interpret failures in visual grounding and object detection tasks, surpassing existing methods across multiple evaluation metrics. The code will be released at https://github.com/RuoyuChen10/VPS.

Paper Structure

This paper contains 27 sections, 1 theorem, 21 equations, 13 figures, 9 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Problem Formulation
  5. Visual Precision Search
  6. Theory Analysis
  7. Experiments
  8. Experimental Setup
  9. Evaluation Metrics
  10. Faithfulness Analysis
  11. Faithfulness on Grounding DINO
  12. Faithfulness on Florence-2
  13. Interpreting Failures in Object-Level Tasks
  14. Interpreting REC Failures
  15. Interpreting Detection Failures
  16. ...and 12 more sections

Key Result

Theorem 1

Consider two sub-sets $S_{A}$ and $S_{B}$ in set $V$, where $S_{A} \subseteq S_{B} \subseteq V$. Given an element $\alpha$, where $\alpha \in V \setminus S_{b}$. Assuming that $\alpha$ is contributing to model interpretation, then, the function $\mathcal{F}(\cdot)$ in Eq. submodular_function satisfi and monotonic non-negative, $\mathcal{F}(S_A \cup \{\alpha\}) - \mathcal{F}(S) \ge 0$, and thus, $\

Figures (13)

  • Figure 1: Illustration of our Visual Precision Search interpretation method, which more precisely identifies key sub-regions in object-level foundation model decision-making compared to gradient-based and perturbation-based methods.
  • Figure 2: Framework of the proposed Visual Precision Search method for interpreting an object-level foundation model. The input is first sparsified into a set of sub-regions and then interpreted across different instances. A submodular function guides the search for significant sub-regions, updating the ordered subset iteratively, and ultimately generating the instance-level attribution map.
  • Figure 3: Visualization results of Grounding DINO for interpreting object localization tasks on the MS COCO, RefCOCO, and LVIS datasets. The first column shows the saliency map with the ground truth box and label. The second column presents detection within the limited search region, and the third columns display the insertion curves.
  • Figure 4: Visualization results of Florence-2 for interpreting object localization tasks on the MS COCO and RefCOCO datasets.
  • Figure 5: Visualization of the method for discovering what causes the Grounding DINO incorrectly grounded on RefCOCO.
  • ...and 8 more figures

Theorems & Definitions (5)

  • Theorem 1: Submodular Properties
  • proof
  • Remark 1: Impact on Sparse Division
  • Remark 2: Applicable Models
  • proof