BioDet: Boosting Industrial Object Detection with Image Preprocessing Strategies

AI summary

Overview

• Research area: Computer vision for robotics — specifically 2D object detection and 6D pose estimation of unseen objects in industrial bin-picking environments, built on BOP-Industrial benchmarks.
• Technical level: Intermediate. The paper assumes familiarity with foundation-model segmentation (SAM, FastSAM), open-vocabulary detection (Grounding-DINO), feature matching (DINOv2 / CNOS / SAM6D), and BOP detection metrics (AP, VSD, MSSD, MSPD).
• Scope: A plug-in image-preprocessing pipeline (low-light enhancement plus open-vocabulary background removal) that improves the detection stage of existing unseen-object pose estimation baselines at negligible extra runtime.

What This Paper Is About

Generalizable model-based pipelines for 6D pose estimation first detect and localize an object in an image, then estimate its pose from the cropped region — and pose accuracy depends heavily on detection quality. In industrial settings (extreme lighting, heavy clutter, reflective metallic parts, stacked objects), foundation-model detectors such as SAM produce thousands of candidate segments per image, and template matching against rendered CAD models fails because of the domain gap between renders and real industrial images, plus background clutter being matched as objects. This paper proposes a preprocessing pipeline that reduces that domain gap and suppresses background segments so that existing state-of-the-art detection baselines become more reliable in industrial scenes.

Key Contributions

1. A standardized, plug-in 2D detection pipeline for unseen objects in industrial settings that is applied on top of current state-of-the-art baselines (CNOS, SAM6D) rather than replacing them.
2. A selective Low-Light Enhancement Module (LLEM) built on the HVI model and its CIDNet color-and-intensity decoupling network, applied only to images whose average pixel intensity falls below a threshold of 50, avoiding overexposure of reflective metallic parts.
3. A background-removal module using Grounding-DINO with the task-adaptive prompt "Parts frame where multiple parts inside it" to produce a single ROI bounding box, crop the RGB and depth inputs, and discard the false positives prevalent in raw SAM output.
4. Extended evaluation on three BOP-Industrial benchmarks (XYZ-IBD, IPD, ITODD-MV) showing detection gains over SAM6D and CNOS, with an ablation isolating the effect of each preprocessing strategy and a runtime analysis.

Main Findings

• Detection improves substantially on cluttered, poorly lit data. Against SAM6D, the full pipeline reaches AP 47.3 on IPD (+17.3) and 27.0 on XYZ-IBD (+1.3) with SAM, and AP 42.2 on IPD (+13.6) and 26.0 on XYZ-IBD (+2.9) with FastSAM, for mean improvements of +6.2 and +5.5 respectively across the three datasets.
• Gains are largest where conditions are worst. The paper attributes the large IPD gains to that dataset's challenging lighting and cluttered environment. ITODD features good lighting and no background clutter, so the baseline is used without preprocessing there (41.9 BOP for the FastSAM configuration, 39.4 for the SAM configuration).
• Background removal is the dominant contributor. In the ablation with FastSAM, the Grounding-DINO module alone (labelled "Ours (w/o LLEM)") reaches AP 40.9 on IPD and 25.4 on XYZ-IBD, versus 28.6 and 23.1 for the baseline. Low-light enhancement alone (labelled "Ours (w/o G-DINO)") adds only +0.2 on IPD and +0.1 on XYZ-IBD, and the paper states it yields virtually no performance gain on its own.
• Note on a reported discrepancy: the ablation text states that using only the GDINO module outperforms the baseline by 13.1% on IPD and 2.9% on XYZ-IBD, while the corresponding Table 1 entries list +12.3 and +2.3. The table values imply the raw AP numbers 40.9 and 25.4.
• Better detection translates into better pose estimation. Feeding the pipeline's segmentation into the SAM6D pose model raises pose AR on XYZ-IBD from 55.0 to 56.2 (+1.2) and on IPD from 40.9 to 61.9 (+21.0), alongside the detection AP increases of +1.3 and +17.3.
• RGB is a better Grounding-DINO input than depth on IPD. With FastSAM, RGB gives AP 42.2 versus 37.4 for the depth modality; with SAM, RGB gives 47.3 versus 42.6. The depth input is produced by masking pixels outside a 150 cm to 200 cm working distance and converting the remainder to a plasma colormap. The paper reports the RGB-over-depth improvement in the text as 4.3% (FastSAM) and 4.4% (SAM), while Table 3's values imply 4.8 and 4.7 — another reported discrepancy.
• The added preprocessing stage is cheap. On a single RTX4090, preprocessing costs 0.176 s (XYZ-IBD, FastSAM) and 0.145 s (IPD, FastSAM), or 0.139 s and 0.120 s with the SAM configuration. The main proposal-and-matching stage is faster than the baseline in every case (for example, 0.428 vs 0.479 s with FastSAM on XYZ-IBD, and 1.943 vs 2.367 s with SAM), because it operates on lower-resolution crops. Total time increases by 0.125 s and 0.091 s for the FastSAM configurations, decreases by 0.285 s for the SAM configuration on XYZ-IBD, and increases by 0.013 s for the SAM configuration on IPD.

Methodology in Plain English

The pipeline keeps the existing detection-and-matching design and inserts two preprocessing steps in front of it.

1. Check the lighting first. Compute the average pixel intensity of the input image. If it is below 50, run the HVI low-light enhancement model (CIDNet); otherwise leave the image untouched. This selectivity avoids overexposing already-bright metallic surfaces.
2. Crop out the busy background. Feed the enhanced RGB image to Grounding-DINO with the text prompt "Parts frame where multiple parts inside it". In a bin-picking scene the parts are usually confined to a container or platform, so this prompt reliably yields one bounding box around the region of interest. That box is used to crop both RGB and depth, with camera intrinsics adjusted to account for the shifted image origin, and the box is stored so detections and poses can be mapped back to the original image. The paper also describes an alternative depth input: mask pixels outside a 150–200 cm working distance range and render the remainder as a plasma colormap before prompting Grounding-DINO.
3. Segment inside the crop. Apply SAM (ViT-H) or FastSAM to the cropped image, prompting with a uniform grid of points, then filter proposals by a confidence threshold and non-maximum suppression to remove redundant masks.
4. Match masks to CAD templates. Render templates by sampling object poses in SE(3) and encode both templates and proposals with a DINOv2 ViT-L backbone, extracting a global class token and local patch features. Following the SAM6D framework, score each proposal against each template along three axes — semantic similarity of global embeddings, appearance similarity of patch features, and geometric consistency of spatial alignment — combine them into a single metric with higher weight on reliable cues such as geometry in visible regions, and assign the highest-scoring template's label and its 6D pose hypothesis to the proposal.

Evaluation uses BOP metrics: detection AP averaged over IoU thresholds from 0.50 to 0.95 in steps of 0.05, and pose AR under VSD, MSSD and MSPD.

Why This Matters

Research impact. The paper reframes object detection, rather than pose refinement, as the bottleneck for unseen-object 6D pose estimation in industrial domains. It shows that a lightweight, model-agnostic preprocessing layer — not a new architecture — can produce double-digit AP gains on hard industrial benchmarks, and that the pipeline is applicable to any existing baseline.

Real-world applications.

• Robotic bin-picking in factories, where parts must be localized and grasped from cluttered, stacked containers.
• Flexible automation lines that need to handle new or frequently changing parts without retraining a detector for each target object.
• Industrial quality inspection and assembly tasks where reflective metallic parts and poor factory lighting degrade vision systems.
• Any vision pipeline that already uses SAM-style segmentation, where a cheap ROI cropping step could remove background false positives.

Industry relevance. The runtime analysis matters commercially: on the tested hardware the preprocessing adds roughly a tenth to a fifth of a second in the FastSAM configurations and the total per-image time stays near or below the baseline in several settings, so the accuracy gains do not come at the cost of impractical inference latency. The depth-versus-RGB finding is also practical, suggesting that on datasets where depth is inaccurate, RGB prompting is the safer configuration.

Future Directions

• Reconciling the reported discrepancies. The ablation percentages in the text differ from the corresponding Table 1 values on IPD and XYZ-IBD, and the RGB-versus-depth percentages in the text differ from Table 3; clarifying these would strengthen trust in the ablation conclusions.
• Extending beyond 2D detection. The paper evaluates detection AP and downstream pose AR, but the pipeline stops at producing detections and hypotheses; how far the end-to-end pose accuracy can be pushed by also improving refinement is left open.
• Reducing prompt dependence. Background removal relies on a hand-written prompt tuned for bin-picking scenes ("Parts frame where multiple parts inside it"); whether the approach generalizes to non-bin-picking industrial layouts or needs per-scene prompt engineering is not reported.
• Improving depth-based ROI localization. The paper finds depth input underperforms RGB on IPD because the depth map is less accurate; whether better depth filtering or calibration would close that gap is not reported.

Target Audience

Robotics and computer vision researchers working on 6D pose estimation, unseen-object detection and bin-picking; engineers integrating foundation models (SAM, FastSAM, Grounding-DINO, DINOv2) into production vision pipelines; and practitioners who want a low-overhead way to make existing detection baselines more robust to poor lighting and background clutter without retraining. Readers seeking architectural novelty will find the contribution is engineering and empirical rather than a new model, so familiarity with the BOP benchmarks and CNOS/SAM6D pipelines is helpful.

Authors’ 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.