ProCal: Inference-Time Proposal Calibration for Open-Vocabulary Object Detection

2026-07-02 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, quick

Summary

ProCal is an inference-time Proposal Calibration method designed to enhance open-vocabulary object detection by improving the localization quality of classification scores from frozen Vision-Language Models (VLMs). While VLMs like CLIPSelf ViT-L/14 are used as detector backbones for unseen categories, their classification scores often lack precise object position and scale recognition. ProCal addresses this by observing that VLMs can distinguish foreground and background regions. It computes a "proposal prior" by combining a localization-aware foreground score, which identifies object areas, and a background-aware suppression score, which measures background resemblance. This process effectively suppresses false novel activations on background proposals and consistently ranks true novel proposals higher. When applied to CLIPSelf ViT-L/14, ProCal achieved an APr improvement of +2.5 on the OV-LVIS benchmark.

Key takeaway

For Machine Learning Engineers developing open-vocabulary object detection systems, integrating ProCal can significantly enhance novel category performance. If your current VLM-based detector struggles with precise localization or false positives on background regions, consider implementing ProCal's inference-time proposal calibration. This approach, which improved APr by +2.5 on OV-LVIS for CLIPSelf ViT-L/14, offers a direct method to refine classification scores and improve overall detection accuracy for unseen objects.

Key insights

ProCal improves open-vocabulary object detection by calibrating VLM classification scores with foreground/background awareness at inference time.

Principles

• VLMs can distinguish foreground/background regions.
• Combining localization and suppression scores refines proposals.
• Inference-time calibration mitigates ranking miscalibration.

Method

ProCal computes a proposal prior by combining a localization-aware foreground score and a background-aware suppression score to improve classification score localization quality.

In practice

• Apply ProCal to VLM-backed object detectors.
• Use foreground/background scores for proposal reranking.
• Enhance novel object detection performance.

Topics

• Open-Vocabulary Object Detection
• Vision-Language Models
• Proposal Calibration
• Inference-Time Optimization
• CLIPSelf ViT-L/14
• OV-LVIS Benchmark

Best for: Research Scientist, AI Scientist, Computer Vision Engineer, Machine Learning Engineer

Related on AIssential

• MemOVCD: Training-Free Open-Vocabulary Change Detection via Cross-Temporal Memory Reasoning and Global-Local Adaptive Rectification — MemOVCD improves open-vocabulary change detection by integrating cross-temporal memory reasoning and global-local adaptive rectification.
• Adaptive Slicing-Assisted Hyper Inference for Enhanced Small Object Detection in High-Resolution Imagery — ASAHI adaptively slices high-resolution images to enhance small object detection, reducing computation and improving performance.
• Compact Object-Level Representations with Open-Vocabulary Understanding for Indoor Visual Relocalization — OpenReLoc uses object-level representations and open-vocabulary semantics for robust indoor visual relocalization.
• Test-Time Training for Robust Text-Guided Open-Vocabulary Object Counting — Dual-TTT enhances text-guided open-vocabulary object counting robustness in degraded conditions by test-time training a lightweight denoising module.
• An Open-Source Benchmark and Baseline for Multi-temporal Referring Segmentation — Multi-temporal Referring Segmentation (MTRS) is a new task and benchmark for language-guided temporal change detection in images.
• SPOT-E: Test-Time Entropy Shaping with Visual Spotlights for Frozen VLMs — SPOT-E enhances VLM performance on evidence-intensive tasks by using test-time entropy shaping with visual spotlights to improve grounding.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.