Detecting 3D Line Segments for 6DoF Pose Estimation with Limited Data (VISAPP 2026)

Deep learning–based method for direct 3D line segment detection in RGB-D data, derived from LETR, together with evaluation code using 3D line segments for 6DoF bin pose estimation.

Installation

Bins Dataset (3D Scans)

We provide the Bins dataset used for training and evaluation.

• Available Modalities and Format:
  • RGB images (.png / .exr)
  • Structured XYZ point clouds stored in .exr
  • Surface normals stored in .exr
  • Simplified CAD models of bins in scans .stl
  • Annotations and split into train, validation and test set provided in two .json files
• Dataset size: ~37 GB
• Download: https://415102.xyz/share/8qK7ZAuR

The dataset contains 3D scans of bins captured from four different sources:

• cam1 — real 3D scanner (with intensity image instead of RGB)
• cam2 — real 3D scanner (with RGB color image)
• cam3 — real 3D scanner (with RGB color image)
• synth — synthetic scans (with RGB color image)

---

Global Statistics

• Total samples: 1707
• Real samples: 1050
• Synthetic samples: 657

Samples by source:

• cam1: 520
• cam2: 265
• cam3: 265
• synth: 657

---

Resolution Statistics (RGB / Intensity)

• cam1: 2064×1544, single-channel intensity (mostly .exr, 3 .png)
• cam2: 1120×800, 3-channel RGB (.exr)
• cam3: 1120×800, 3-channel RGB (.exr)
• synth: 2064×1544, 3-channel RGB (.png)

---

Dataset Split

• Train: 1507 samples

  • Real: 850
  • Synthetic: 657

• Validation: 100 samples

  • Real: 100

• Test: 100 samples

  • Real: 100

Note: Both validation and test splits contain only real samples.

---

Bin CAD Models

• Located in: <dataset_root>bin_models/
• Includes:
  • One high-quality scanned 3D model
  • Multiple simplified CAD models of bins (.stl)

---

Annotations

Each annotation entry (files train_val.json and test.json in dataset root) contains:

• Transformation of bin model (transform)
• Bin model identifier (bin_model_id)
• Additional metadata:
  • device
  • is_synthetic
  • bin_height
  • associated file names for sample

Experiments

Training models to test the effect of query count:

sh experiments/bins_3D/query_count_experiment.sh \
    --bins_path /path/to/dataset/train_val.json \
    --bins_lines_annotation_dir experiments/bins_3D/lines_annotation/outer_edge

Training models with and without synthetic samples to evaluate their contribution:

sh experiments/bins_3D/synth_no_synth_experiment.sh \
    --bins_path /path/to/dataset/train_val.json \
    --bins_lines_annotation_dir experiments/bins_3D/lines_annotation/outer_edge

Training models to evaluate the effect of cutout augmentation:

sh experiments/bins_3D/cutout_experiment.sh \
    --bins_path /path/to/dataset/train_val.json \
    --bins_lines_annotation_dir experiments/bins_3D/lines_annotation/outer_edge

Training the final model based on the results of the previous ablation studies:

sh experiments/bins_3D/train_final.sh \
    --bins_path /path/to/dataset/train_val.json \
    --bins_lines_annotation_dir experiments/bins_3D/lines_annotation/outer_edge

Training

Evaluation

To evaluate trained model run inference script infer_dataset.py, which for every sample produce prediction file. :

python src/infer_dataset.py \
    --dataset_name bins \
    --model /path/to/model.pth \
    --output_directory predictions \
    --split test \
    --bins_path /path/to/dataset/test.json \
    --bins_lines_annotation_dir experiments/bins_3D/lines_annotation/outer_edge \
    --bins_no_preload \
    --bins_input_width 516 \
    --bins_input_height 386

Citation

@conference{visapp26_line_pose,
author={Matej Mok and Lukáš Gajdošech and Michal Mesároš and Martin Madaras and Viktor Kocur},
title={Detecting 3D Line Segments for 6DoF Pose Estimation with Limited Data},
booktitle={Proceedings of the 21st International Conference on Computer Vision Theory and Applications - Volume 3: VISAPP},
year={2026},
pages={616-623},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0014462600004084},
isbn={978-989-758-804-4},
}

Acknowledgements

This code is based on the implementations of **LETR: Line Segment Detection Using Transformers without Edges **.