Abstract
Task-generic promptable segmentation aims to identify and segment objects based on high-level linguistic descriptions (e.g., "find the polyp"). However, Vision-Language Models (VLMs) often suffer from hallucinations in complex scenes, leading to inaccurate prompts for downstream segmentation models. We present INT (Instance-specific Negative Mining), a training-free, test-time adaptation method that refines instance-specific prompts by measuring VLM confidence drops upon object occlusion. By iteratively "mining" its own errors, INT outperforms several supervised benchmarks on challenging camouflaged and medical datasets.
Introduction
Traditional image segmentation relies heavily on large-scale manual labeling. While models like the Segment Anything Model (SAM) have introduced "promptable" segmentation, they typically require per-image manual intervention such as points or boxes. Task-generic segmentation seeks to automate this by using a single high-level prompt for an entire dataset.
The primary challenge in this paradigm is the reliability of the VLM. In complex scenarios—such as a frog camouflaged against tree bark—VLMs may misidentify background elements as target objects. Without ground-truth labels at test time, these errors propagate, resulting in poor segmentation masks. INT addresses this by using the model's own sensitivity to occlusion as a self-correction mechanism.
Methodology: Instance-Specific Negative Mining
INT treats the VLM's internal consistency as a proxy for ground truth. The pipeline operates in two primary stages:
1. Negative Mining for Prompt Refinement
The core innovation of INT is using image inpainting to verify VLM predictions.
- Candidate Generation: A VLM identifies potential objects within image patches.
- Negative Mining: An inpainting module (e.g., Stable Diffusion) "erases" the suspected object. The system then compares the VLM's confidence score before and after erasure.
- The Logic: A significant drop in confidence indicates a high-probability match, whereas a minimal change suggests the VLM was attending to irrelevant background features (a "hard false" negative).
This process is repeated iteratively. By cumulatively multiplying these "difference scores" over several epochs, the model amplifies consistent detections and suppresses sporadic hallucinations.
2. Semantic Mask Generation
Refined prompts are passed to GroundingDINO to generate bounding boxes, which then guide SAM to produce precise masks. A final CLIP-based semantic similarity check filters out residual noise, ensuring the mask aligns with the original task description.
Experiments and Results
INT was evaluated on six datasets across two domains: Camouflaged Object Detection (COD10K, CAMO, CHAMELEON) and Medical Image Segmentation (CVC-ColonDB, Kvasir-SEG, ISIC).
Performance Benchmarks
INT demonstrated superior performance, often exceeding supervised methods:
- COD10K: INT achieved a Mean Absolute Error (M) of 0.037, surpassing scribble-supervised (0.038) and point-supervised (0.039) models.
- Medical Domain: On CVC-ColonDB, INT reached an E-measure (Eφ) of 0.589, significantly outperforming other task-generic methods like GenSAM (0.494).
The iterative refinement typically converges within five iterations, effectively "cleaning" the prompts without requiring domain-specific fine-tuning.
Discussion and Conclusion
INT represents a significant step toward autonomous, high-precision segmentation. By leveraging the internal consistency of frozen VLMs through occlusion, it achieves state-of-the-art results in a training-free manner.
However, the computational cost of running multiple models (VLM, Diffusion, GroundingDINO, SAM) iteratively is high, limiting real-time applicability. Future research may focus on distilling this negative mining logic into more efficient, single-pass architectures or extending the method to more abstract task descriptions.
Summary updated