PRISMA

PRISMA is a computational photography pipeline that performs multiple inferences (refere as "bands") from any image or video. Like light pasing through a prism that bends it into different wavelengths, this pipeline expands images into data that can be use for 3D reconstruction or realtime post-processing operations.

It's a combination of different algorithms and open sourced pre-train models such as:

• Monocular depth (MiDAS v3.1, ZoeDepth, Marigold, PatchFusion, Depth_Anything)
• Optical flow (RAFT, GMFlow)
• Segmentation mask (mmdet)
• camera pose (COLMAP)

The resulting bands are stored in a folder with the same name as the input file. Each band is stored as a single .png or .mp4 file. And can be imported on:

• Estimated depth can be importer to Blender projects using this blender project, also COLMAP scenes can be imported using this addon
• GlslViewer for applying real-time shaders
• Videos can be use for both NeRFs (like NVidia's Instant-ngp) or Gaussian Splatting training.

Notes:

• Infered depth is exported by default as a heatmap that can be decoded realtime using LYGIA's heatmap GLSL/HLSL sampling.
• optical flow is encoded as HUE (angle) and saturation which also can be decoded realtime using LYGIA opticalFlow GLSL/HLSL sampler.

Install

Main dependencies:

• Conda
• COLMAP

git clone git@github.com:patriciogonzalezvivo/prisma.git
cd prisma

conda env create -f environment.yml
conda activate prisma
sh download_models.sh

# Install mmcv (for mmdetection)
pip install -U openmim
mim install mmengine
mim install "mmcv-full==1.7.1"

How it works?

a. Process

We start by processing an image or video. Let's start by processing an image:

python process.py -i data/gog.jpg

With out providing an --output this will create a folder with the same filename which will contain all the derived bands (rgba, flow, mask and depth_*).

gog.jpg
gog/
├── depth_patchfusion.png
├── mask.png
├── metadata.json
└── rgba.png

In the forlder you will find a metadata.json file that contains all the metadata associated with the original image or video.

{
    "bands": {
        "rgba": {
            "url": "rgba.png"
        },
        "depth_patchfusion": {
            "url": "depth_patchfusion.png",
            "values": {
                "min": {
                    "value": 1.6147574186325073,
                    "type": "float"
                },
                "max": {
                    "value": 11.678544044494629,
                    "type": "float"
                }
            }
        },
        "mask": {
            "url": "mask.png",
            "ids": [
                "person",
                "bird",
                "cat",
                "dog",
                "horse",
                "sheep",
                "cow",
                "elephant",
                "bear",
                "zebra",
                "giraffe"
            ]
        }
    },
    "width": 934,
    "height": 440,
    "principal_point": [
        467.0,
        220.0
    ],
    "focal_length": 641.0616195031489,
    "field_of_view": 37.88246641919117
}

Currently PRISMA supports multiple depth estimation algorithms. You can select which one to use by providing the --depth|-d argument: depth_midas, depth_zoedepth, depth_patchfusion, depth_marigold or all. By defualt images will be processed using depth_patchfusion, while videos will use depth_anything.

When processing videos, by default PRISMA creates the least ammount of data by creating a single .png or .mp4 for each band. In the case of videos data like min/max values will be stored on .cvs.

it's possible to save extra data by setting the --extra|-e level number.

0. store bands as a single .png and .mp4 (video have usually an associated .csv file)
1. store images as .ply point clouds, for videos it extracts the reslting frames as .png
2. store optical flow from videos as .flo files.
3. store inferenced depth as .npy files.

Let's try now extracting all depth models and individual frames from a video:

python process.py -i data/rocky.mp4 -d all -e 1

Which produce the folowing folder structure:

rocky.mp4
rocky/
├── depth_anything/
│   ├── 000000.png
│   ├── 000001.png
│   ├── ...
│   └── 000110.png
├── depth_anything_max.csv
├── depth_anything_min.csv
├── depth_anything.mp4
├── depth_marigold/
│   ├── 000000.png
│   ├── 000001.png
│   ├── ...
│   └── 000110.png
├── depth_marigold_max.csv
├── depth_marigold_min.csv
├── depth_marigold.mp4
├── depth_midas/
│   ├── 000000.png
│   ├── 000001.png
│   ├── ...
│   └── 000110.png
├── depth_midas_max.csv
├── depth_midas_min.csv
├── depth_midas.mp4
├── depth_patchfusion/
│   ├── 000000.png
│   ├── 000001.png
│   ├── ...
│   └── 000110.png
├── depth_patchfusion_max.csv
├── depth_patchfusion_min.csv
├── depth_patchfusion.mp4
├── depth_zoedepth/
│   ├── 000000.png
│   ├── 000001.png
│   ├── ...
│   └── 000110.png
├── depth_zoedepth_max.csv
├── depth_zoedepth_min.csv
├── depth_zoedepth.mp4
├── flow_raft/
│   ├── 000000.png
│   ├── 000001.png
│   ├── ...
│   └── 000110.png
├── flow_raft.csv
├── flow_raft.mp4
├── flow_gmflow/
│   ├── 000000.png
│   ├── 000001.png
│   ├── ...
│   └── 000110.png
├── flow_gmflow.csv
├── flow_gmflow.mp4
├── images/
│   ├── 000000.png
│   ├── 000001.png
│   ├── ...
│   └── 000110.png
├── mask/
|   ├── 000000.png
|   ├── 000001.png
|   ├── ...
|   └── 000110.png
├── mask.mp4
|── sparse/
|   └── 0/
|       ├── cameras.bin
|       ├── images.bin
|       ├── points3D.bin
|       └── points3D.txt
|── camera_pose.csv
|── colmap.db
├── metadata.json
└── rgba.mp4

b. Visualize

View the resulting bands from the processed image/video using ReRun:

```bash
python view.py -i data/rocky

c. Concatenate bands

In order to export the bands as a single image or video you can use the concat.py script:

python concat.py -i data/gog -o test.png

Licenses and Credits

This pipeline is Copyright (c) 2024, Patricio Gonzalez Vivo and Licensed under CC BY-NC-SA 4.0 please reach out to patriciogonzalezvivo at gmail dot com, for getting a comercial license.

All the models and software used by it are commercial ready licenses like MIT, Apache and BSD.

Depth estimation (MiDAS 3.1)

Paper: Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-shot Cross-dataset Transfer

License: MIT

Code Repo: isl-org/MiDaS

Use:

depth_midas.py --input <IMAGE/VIDEO> --output <IMAGE/VIDEO>

Citation:

@ARTICLE {Ranftl2022,
    author  = "Ren\'{e} Ranftl and Katrin Lasinger and David Hafner and Konrad Schindler and Vladlen Koltun",
    title   = "Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-Shot Cross-Dataset Transfer",
    journal = "IEEE Transactions on Pattern Analysis and Machine Intelligence",
    year    = "2022",
    volume  = "44",
    number  = "3"
}

Citation for DPT-based model:

@article{Ranftl2021,
    author    = {Ren\'{e} Ranftl and Alexey Bochkovskiy and Vladlen Koltun},
    title     = {Vision Transformers for Dense Prediction},
    journal   = {ArXiv preprint},
    year      = {2021},
}

Depth Estimation (ZoeDepth)

Paper: Zero-shot Transfer by Combining Relative and Metric Depth

License: MIT

Code Repo: isl-org/ZoeDepth

Use:

depth_zoedepth.py --input <IMAGE/VIDEO> --output <IMAGE/VIDEO>

Citation

@misc{https://doi.org/10.48550/arxiv.2302.12288,
    doi = {10.48550/ARXIV.2302.12288},
    url = {https://arxiv.org/abs/2302.12288},
    author = {Bhat, Shariq Farooq and Birkl, Reiner and Wofk, Diana and Wonka, Peter and Müller, Matthias},  
    keywords = {Computer Vision and Pattern Recognition (cs.CV), FOS: Computer and information sciences, FOS: Computer and information sciences},
    title = {ZoeDepth: Zero-shot Transfer by Combining Relative and Metric Depth},  
    publisher = {arXiv},
    year = {2023},
    copyright = {arXiv.org perpetual, non-exclusive license}
}

Depth Estimation (PatchFusion)

Paper: PatchFusion: An End-to-End Tile-Based Framework for High-Resolution Monocular Metric Depth Estimation

License: MIT

Code Repo: zhyever/PatchFusion

Use:

depth_patchfusion.py --input <IMAGE/VIDEO> --output <IMAGE/VIDEO>

Note: This pretrained model needs to be downloaded and placed in the models/ folder.

Citation

@article{li2023patchfusion,
    title={PatchFusion: An End-to-End Tile-Based Framework for High-Resolution Monocular Metric Depth Estimation}, 
    author={Zhenyu Li and Shariq Farooq Bhat and Peter Wonka},
    year={2023},
    eprint={2312.02284},
    archivePrefix={arXiv},
    primaryClass={cs.CV}}

Depth Estimation (Marigold)

Paper: Marigold: Repurposing Diffusion-Based Image Generators for Monocular Depth Estimation

License: Apache

Code Repo: prs-eth/Marigold

Use:

depth_marigold.py --input <IMAGE/VIDEO> --output <IMAGE/VIDEO>

Citation

@misc{ke2023repurposing,
      title={Repurposing Diffusion-Based Image Generators for Monocular Depth Estimation}, 
      author={Bingxin Ke and Anton Obukhov and Shengyu Huang and Nando Metzger and Rodrigo Caye Daudt and Konrad Schindler},
      year={2023},
      eprint={2312.02145},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Depth Estimation (Depth Anything)

Paper: Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data

License: Apache

Code Repo: LiheYoung/Depth-Anything

Use:

depth_anything.py --input <IMAGE/VIDEO> --output <IMAGE/VIDEO>

Citation

@article{depthanything,
      title={Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data}, 
      author={Yang, Lihe and Kang, Bingyi and Huang, Zilong and Xu, Xiaogang and Feng, Jiashi and Zhao, Hengshuang},
      journal={arXiv:2401.10891},
      year={2024}
}

Optical Flow (RAFT)

Based on https://github.com/SharifElfouly/opical-flow-estimation-with-RAFT

Seems to be very good: Optical Flow Estimation Benchmark

Paper: RAFT: Recurrent All Pairs Field Transforms for Optical Flow

License: BSD

Code Repo: princeton-vl/RAFT

Use:

flow_raft.py --input <IMAGE/VIDEO> --output <IMAGE/VIDEO>

Optical Flow (GMFlow)

Paper: GMFlow: Learning Optical Flow via Global Matching

License: Apache

Code Repo: haofeixu/gmflow

Use:

flow_gmflow.py --input <IMAGE/VIDEO> --output <IMAGE/VIDEO>

Segmentation (MMDetection)

Code Repo: MMDetection

License: Apache

Use:

mask_mmdet.py --input <IMAGE/VIDEO> --output <IMAGE/VIDEO>

Citation:

@article{mmdetection,
    title   = {{MMDetection}: Open MMLab Detection Toolbox and Benchmark},
    author  = {Chen, Kai and Wang, Jiaqi and Pang, Jiangmiao and Cao, Yuhang and
              Xiong, Yu and Li, Xiaoxiao and Sun, Shuyang and Feng, Wansen and
              Liu, Ziwei and Xu, Jiarui and Zhang, Zheng and Cheng, Dazhi and
              Zhu, Chenchen and Cheng, Tianheng and Zhao, Qijie and Li, Buyu and
              Lu, Xin and Zhu, Rui and Wu, Yue and Dai, Jifeng and Wang, Jingdong
              and Shi, Jianping and Ouyang, Wanli and Loy, Chen Change and Lin, Dahua},
    journal= {arXiv preprint arXiv:1906.07155},
    year={2019}
}