Research map / Detection and segmentation / Open-set and few-shot detection

An Application-Agnostic Automatic Target Recognition System Using Vision Language Models

AAAI 2025 · Detection and segmentation › Open-set and few-shot detection · approach: Open-vocabulary object detection · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Builds an application-agnostic automatic target recognition system on a pre-trained open-vocabulary detector (MM-OVOD), letting users define target classes via text descriptions and/or image exemplars, with sequential bounding-box matching, tubelet linking, and mosaic/KDE visualization, evaluated on UXO detection on airfield runways.

Problem

Traditional ATR systems require retraining deep detectors to add new classes, which takes days or weeks, so they cannot quickly adapt to new or nuanced targets (e.g., unexploded ordnance) in the field.

Approach

Use an open-vocabulary detector whose classes are defined at runtime by natural-language descriptions and/or image exemplars, combined with post-processing across overlapping video frames (tubelet matching with box re-scoring, tubelet linking) and a stitched mosaic with kernel density heatmap to aggregate detections over time and multiple drone passes.

Key results

• MM-OVOD with natural-language class descriptions achieved weighted AP of 0.75 and weighted F1 of 0.69 on 410 UXO objects across 21 runways, despite no training on the novel UXO classes.
• Per-class F1 ranged from 0.43 (155MM) to 0.92 (BLU63) at confidence threshold 0.1, with overall accuracy 0.68.
• Probability of correct classification stays above 60% at low confidence thresholds, per the paper's classification ROC results.

Measured effects reported in the paper

Surprising findings

• The 155MM shell performed worst (F1 0.43, recall 0.28) because its text descriptions overlapped with other cylindrical UXO classes.
• Tubelet linking did not help in this application and filtering short tubelets reduced peak detection probability, so the authors dropped it.

Limitations

• False alarm density is overestimated because mosaic alignment for de-duplicating overlapping passes was not yet applied.
• Performance is weaker for classes with few distinguishing visible features, such as 155MM, and for BLU26 which is sometimes not detected.
• Evaluation is limited to a single application (UXO on airfield runways) with a proprietary government test dataset.

Experimental setup

Datasets: ONR RLAC UXO test dataset (21 runways, 410 UXO objects)

Benchmarks: ONR RLAC UXO test and evaluation dataset

Metrics: probability of detection false alarm density probability of correct classification precision recall F1-score weighted AP

Baselines: DITO OV-DQUO LaMI-DETR YOLO-World

Builds on: MM-OVOD Seq-BBox-Matching CenterNet2 Seq-NMS

Closest papers by meaning

• Revisiting Few-Shot Object Detection with Vision-Language Models (NeurIPS 2024 · similarity 0.83)
• Simple Image-Level Classification Improves Open-Vocabulary Object Detection (AAAI 2024 · similarity 0.83)
• Roboflow100-VL: A Multi-Domain Object Detection Benchmark for Vision-Language Models (NeurIPS 2025 · similarity 0.82)
• The Devil is in the Fine-Grained Details: Evaluating Open-Vocabulary Object Detectors for Fine-Grained Understanding (CVPR 2024 · similarity 0.82)
• Taming Self-Training for Open-Vocabulary Object Detection (CVPR 2024 · similarity 0.82)
• How to Evaluate the Generalization of Detection? A Benchmark for Comprehensive Open-Vocabulary Detection (AAAI 2024 · similarity 0.82)