BioDet: Boosting Industrial Object Detection with Image Preprocessing Strategies

TL;DR

This work targets the critical detection bottleneck in industrial 6D pose estimation by introducing a plug-in 2D unseen-object detection pipeline. It combines a low-light enhancement module (LLEM) with background removal via Grounding-DINO, followed by SAM-based segmentation and multi-faceted template matching using DINOv2 features to robustly assign CAD IDs in $SE(3)$ pose space. The approach reduces domain gaps and suppresses background clutter, yielding significant improvements in detection AP and pose AR on challenging industrial benchmarks (XYZ-IBD, IPD, ITODD-MV) with only modest overhead. Practically, the method enhances robustness under extreme lighting and clutter, making industrial automation more flexible and reliable when detecting unseen objects using existing baselines.

Abstract

Accurate 6D pose estimation is essential for robotic manipulation in industrial environments. Existing pipelines typically rely on off-the-shelf object detectors followed by cropping and pose refinement, but their performance degrades under challenging conditions such as clutter, poor lighting, and complex backgrounds, making detection the critical bottleneck. In this work, we introduce a standardized and plug-in pipeline for 2D detection of unseen objects in industrial settings. Based on current SOTA baselines, our approach reduces domain shift and background artifacts through low-light image enhancement and background removal guided by open-vocabulary detection with foundation models. This design suppresses the false positives prevalent in raw SAM outputs, yielding more reliable detections for downstream pose estimation. Extensive experiments on real-world industrial bin-picking benchmarks from BOP demonstrate that our method significantly boosts detection accuracy while incurring negligible inference overhead, showing the effectiveness and practicality of the proposed method.

Figures (4)

• Figure 1: (Left) Overview of our pipeline. Images from industrial datasets Kalra_2024_CVPRhuang2025xyzDrost_2017_ICCV are first enhanced by our Low-Light Enhancement Module (LLEM) yan2025hvi, then processed by Grounding-DINO liu2023grounding for ROI detection; the highest-confidence crop is passed to SAM6D lin2023sam6d for final segmentation. (Right) ROI detection example. Either a plasma-mapped depth image or the raw RGB input is prompted with “Parts frame where multiple parts are inside it” to generate a bounding box for background removal.
• Figure 2: Qualitative results on the XYZ–IBD huang2025xyz (top row) and IPD Kalra_2024_CVPR (bottom row) datasets. Compared to the SOTA baseline method SAM6D lin2023sam6d, our detection pipeline exhibits fewer false positive detections and more true positives, demonstrating the effectiveness of our image enhancement and background removal modules under such industrial bin-picking setups.
• Figure 3: Ablation study of the low-light enhancement model (LLEM). Row 1 shows GDINO without LLEM, and row 2 shows GDINO with LLEM.
• Figure 4: Qualitative comparison of raw SAM segmentation and GDINO-based ROI filtering. Top: results on the IPD dataset Kalra_2024_CVPR. Bottom: results on XYZ-IBD huang2025xyz. GDINO effectively removes background clutter using text prompts, producing cleaner regions of interest compared to raw SAM outputs.