Multi-View

DTU

Issues:

1. Directory Structure
2. SampleSet
3. Read Depth Map

Directory Structure

1. Problems:

   1. The directory structure of the DTU dataset

2. Supports:

   1. Download page: MVS Data Set - 2014 | DTU Robot Image Data Sets

   2. 124 scans - Rectified.zip (123 GB)

      jsmind

      {{{

      }}}

2. SampleSet.zip

   • Supports:

     1. Scan1 and scan6

   • Actions:

     Directory tree

     jsmind

     {{{

     }}}

3. Read Depth Map

Subsets

Issues:

1. mvs_training

Notes:

1. mvs_training

   • Supports:

     1. The DTU dataset is processed for MVSNet by Yaoyao

     2. Image size (dimension recognized by identify): 640x512

     3. The focal length of the camera matrix inside the train/ folder corresponds to the camera parameters of feature maps. Focal length get scaled alongside only downsampling operations.

        flowchart LR A["Rectified image
        (1600x1200)"] --> B["Downsample
        (800x600)"] --> C["Crop
        (640x512)"] --> D["Feature map
        (160x128)"]

        The principle point (c_x, c_y) get affected by both downsampling and cropping operations.

        flowchart LR A["Princple point
        (823.205, 619.071)"] --> B["Download
        (411.6, 309.54)"] --> C["Crop
        (331.6, 265.54)"] --> D["Feature map
        (82.9, 66.38)"]
        Table

        {{{

        SampleSet	Decupled	train
        SampleSet/MVS Data/Calibration/cal18$ cat pos_001.txt 2607.429996 -3.844898 1498.178098 -533936.661373 -192.076910 2862.552532 681.798177 23434.686572 -0.241605 -0.030951 0.969881 22.540121	mvs_training/dtu/Cameras$ cat 00000000_cam.txt extrinsic 0.970263 0.00747983 0.241939 -191.02 -0.0147429 0.999493 0.0282234 3.28832 -0.241605 -0.030951 0.969881 22.5401 0.0 0.0 0.0 1.0 intrinsic 2892.33 0 823.205 0 2883.18 619.071 0 0 1 425 2.5	mvs_training/dtu/Cameras/train$ cat 00000000_cam.txt extrinsic 0.970263 0.00747983 0.241939 -191.02 -0.0147429 0.999493 0.0282234 3.28832 -0.241605 -0.030951 0.969881 22.5401 0.0 0.0 0.0 1.0 intrinsic 361.54125 0.0 82.900625 0.0 360.3975 66.383875 0.0 0.0 1.0 425.0 2.5

        }}}

   • Actions:

Convert DTU to PCN

Problems:

1. Reorgnize the DTU dataset to the format of the PCN dataset

References:

1. Gemini 2.5P | DTU to PCN Dataset Conversion (May08,25)

Compl. & Acc. (Python)

1. Problems:

   1. The official evaluation for DTU dataset metrics (completeness and accuracy) is performed using MATLAB scripts. However, I don’t have a Matlab license.

   2. I remember there are alternative methods for measurement, which can be implemented in Python.

2. Supports:

   1. Python Implementation ^r1-Fast

   ::: aside

   • References: {{{
     1. Gwencong/Fast-DTU-Evaluation
        Searched by DTU dataset completeness accuracy compute at DDG
     2. Gemini 2.5P - PyTorch CUDA Version Mismatch Error
     3. SO - ImportError: cannot import name ‘packaging’ from ‘pkg_resources’ when …
        Searched by .venv/lib/python3.10/site-packages/torch/utils/cpp_extension.py", line 25, in <module> from pkg_resources import packagin at DDG
     4. unable to build from source - `cannot import name ‘packaging’ from ‘pkg … | GitHub issues
     5. ImportError: cannot import name ‘packaging’ from ‘pkg_resources’’ - CSDN博客
        Searched by ImportError: cannot import name 'packaging' from 'pkg_resources' at DDG }}} :::

3. Actions:

   1. Download source code

      1 2	git clone https://github.com/Gwencong/Fast-DTU-Evaluation.git cd Fast-DTU-Evaluation

   2. Specify dependencies versions in requirements.txt

      1 2 3 4 5 6 7

      --extra-index-url https://download.pytorch.org/whl/cu113 torch==1.12.1+cu113 torchvision==0.13.1+cu113 # torchaudio==0.12.1 # setuptools>=59.6.0,<70.0.0 setuptools==69.5.1 numpy<2

      • Match torch version with the system CUDA version (11.3) ^r2-Gemini

      • torch 1.10 doesn’t support Python higher than 3.10 ^r2-Gemini

      • setuptools==69.5.1 avoids error: ^{r3-SO,r4-Issue,r5-CSDN}

        1	ImportError: cannot import name 'packaging' from 'pkg_resources'

        Traceback

        {{{

        1 2 3 4 5 6 7

        (Fast-DTU-Evaluation) zichen@zichen-X570-AORUS-PRO-WIFI:~/Projects/Fast-DTU-Evaluation$ python chamfer3D/setup.py install Traceback (most recent call last): File "/home/zichen/Projects/Fast-DTU-Evaluation/chamfer3D/setup.py", line 2, in <module> from torch.utils.cpp_extension import BuildExtension, CUDAExtension File "/home/zichen/Projects/Fast-DTU-Evaluation/.venv/lib/python3.10/site-packages/torch/utils/cpp_extension.py", line 25, in <module> from pkg_resources import packaging # type: ignore[attr-defined] ImportError: cannot import name 'packaging' from 'pkg_resources' (/home/zichen/Projects/Fast-DTU-Evaluation/.venv/lib/python3.10/site-packages/pkg_resources/init.py)

        }}}

      • setuptools < 70.0.0 is not in the “PyTorch package index”, which is looked through by uv. So adding unsafe-best-match to allow uv find the compatible version from all souce.

   3. Create environment

      1 2 3 4 5 6 7 8 9 10 11 12

      uv venv -p python3.10 source .venv/bin/activate # Enable PyPI repository for setuptools uv pip install -r requirements.txt --index-strategy unsafe-best-match # Specify compiler versions for CUDA export CC=/usr/bin/gcc-9 export CXX=/usr/bin/g++-9 export CUDAHOSTCXX=/usr/bin/g++-9 cd chamfer3D && python setup.py install # build and install chamfer3D package

      • CUDA 11.3 requires g++ <= 10.0.0 ^r6-

        Example error

        {{{

        1	RuntimeError: The current installed version of x86_64-linux-gnu-g++ (11.4.0) is greater than the maximum required version by CUDA 11.3 (10.0.0). Please make sure to use an adequate version of x86_64-linux-gnu-g++ (>=5.0.0, <=10.0.0).

        }}}

      • update-alternative --config g++ alone for specifying g++-9 doesn’t work (still detect g++-10 (10.5.0)); set CXX explicitly.

   4. Run evaluation with directories of predicted and ground-truth point clouds specified:

      1 2 3

      CUDA_VISIBLE_DEVICES=0 python eval_dtu.py --method scan \ --pred_dir "/mnt/Seagate4T/04-Projects/CasMVSNet_pl-comments/output/no_densify_250929_2views_iter30K_640x512_noDepthReg/combined_ply" \ --gt_dir "/mnt/Seagate4T/05-DataBank/SampleSet/MVS Data" --save --num_workers 1

      • --method constitutes the .ply filename.
      • Results are saved as result.txt under --pred_dir.

Eval Framework Understand

1. Problems:

   1. The Matlab code understanding

2. Supports:

   1. AI explanation ^r1-

   ::: aside

   • References: {{{
     1. DTU Evaluation Framework | caoPhoenix/CasMVSNet | DeepWiki
        Searched by DTU dataset completeness accuracy compute at DDG
     2. DTU Dataset Evaluation | hbb1/2d-gaussian-splatting | DeepWiki }}} :::

Real Estate 10K

Tanks and Temples

References:

1. Tanks and Temples Benchmark | Homepage

Novel View Synthesis

Static

Issues:

1. NeRF
2. MipNeRF360

Notes:

1. NeRF

   • Supports:

     1. Meta info:

        • 8 Feed-forward scenes

Dynamic

Issues:

1. D-NeRF

Notes:

1. D-NeRF:

Point Cloud

ShapeNet

PCN

Issues:

1. Directory Structure
2. Read As PyTorch Datasets
3. O3D Visualization Point Cloud

References:

1. Gemini Deep Research | PCN 点云数据集介绍
2. 元宝 | Python类结构概览方法
3. 元宝 | 笔记转写
4. Visualize point cloud | Open3D 0.19.0 documentation

Notes:

1. Directory Structure

   • Supports:

     1. Point Completion Network

   • Actions:

     (2025-04-27T22:37)

     1. PCN dataset directory structure (~/Downloads/Datasets_Unpack/ShapeNetCompletion)

        1 2 3 4 5 6 7 8 9 10 11

        ~/Downloads/Datasets_Unpack/ShapeNetCompletion$ tree -L 2 . ├── test │ ├── complete │ └── partial ├── train │ ├── complete │ └── partial └── val ├── complete └── partial

     2. In the partial folder, each “complete” point cloud is split into 8 portions.

(2025-04-28T22:30)

2. Read the PCN Dataset as a PyTorch Dataset

   • Supports:

     1. Souce code: PointAttN/dataset.py

   • Actions:

     1. Call graph of the Dataset PCN_pcd^{r2-元宝, r3-元宝}

(2025-05-10T13:49)

3. Visualize Point Clouds using O3D

   • Supports:

     1. o3d read .pcd data^r4-Docs:

        1 2 3 4 5 6 7 8 9 10

        import open3d as o3d pcd = o3d.io.read_point_cloud("/home/zichen/Downloads/Datasets_Unpack/ShapeNetCompletion/train/complete/02691156/10155655850468db78d106ce0a280f87.pcd") vis = o3d.visualization.Visualizer() vis.create_window(width=400, height=400) ctr = vis.get_view_control() param = o3d.io.read_pinhole_camera_parameters("PCN_Cam_Parm_save.json") vis.add_geometry(pcd) ctr.convert_from_pinhole_camera_parameters(param) vis.run()

   • Actions:

     

3D Models

Objverse