adam99goat

MonoPCC: Estimate Monocular Depth in Endoscopic Images

MonoPCC is a PyTorch method for self-supervised monocular depth estimation on endoscopic images, using a photometric-invariant cycle constraint. It is aimed at researchers reproducing or extending depth and pose experiments on surgical-video datasets.

View on GitHub Compare
Stars
23
Forks
1
Stars, last 30 days
+1

Our take

Health score

30/100

How it is scored
Activity
30
Community
15
Issues
53
Pull requests
20

+1 stars in the last 30 days

  • Fewer contributors than 86% of the projects we track

Project history

Daily GitHub snapshots since Oct 7, 2026. History covers stars and forks; health signals are not tracked historically.

Selected snapshot

Stars

24

Forks

1

Change since Nov 14, 2026: 0 stars·0 forks

Overview

MonoPCC is a research implementation for estimating depth from monocular endoscopic images. Its photometric-invariant cycle constraint is intended to address the challenges of learning depth from endoscopic video, where appearance changes can make photometric consistency less reliable.

The repository includes training and evaluation workflows for endoscopic datasets, comparisons with other methods, and experiments on KITTI. It is most relevant to researchers working on surgical-scene reconstruction or self-supervised depth estimation, rather than users seeking a ready-made clinical or production system.

Key Features

  • PyTorch implementation of MonoPCC for monocular depth estimation.
  • Training workflow for the SCARED dataset, initialized from a warm-up checkpoint.
  • Depth evaluation scripts for SCARED and KITTI.
  • Reports comparisons against established monocular depth methods on both datasets.
  • Extends the PCC strategy to EndoDAC and provides reported comparison results.
  • Includes a pose-estimation evaluation workflow and a 3D reconstruction demonstration using estimated depth.
  • Links to model files and prediction results for reproducing reported evaluations.

Use Cases

  • Researchers evaluating self-supervised monocular depth methods on endoscopic video.
  • Surgical-imaging teams prototyping depth-based 3D reconstruction from SCARED sequences.
  • Developers comparing a cycle-constraint approach with other depth-estimation baselines on KITTI.
  • Researchers investigating whether PCC can be integrated with a different depth-estimation backbone, such as EndoDAC.

Getting Started

The README recommends a Conda environment with Python 3.7 and lists the required dependencies. For example, it specifies:

pip3 install torch==1.9.0+cu111 torchvision==0.10.0+cu111 torchaudio==0.9.0

Training requires prepared dataset files and a warm-up checkpoint containing the encoder, depth, and pose model weights. See the README for complete setup, data preparation, training, and evaluation instructions.

Alternatives

  • vggt: VGGT predicts depth and other 3D scene properties with a feed-forward model, rather than learning monocular depth from endoscopic video using a photometric-invariant cycle constraint.
ProjectLanguageLicenseStarsStatus
MonoPCCPythonApache-2.023Not checked yet
vggtPythonOther14.5kQuiet

Considerations

  • This is a research codebase with dataset-specific preparation and evaluation steps, not a packaged end-user application.
  • The documented setup uses older versions of Python, PyTorch, and related libraries. Compatibility with newer environments is not established by the provided information.
  • The example commands use CUDA, and training or evaluation may require a compatible GPU environment.
  • SCARED access requires an additional application, as noted in the README. Reproducing results also depends on preparing the datasets and obtaining the linked model files.

Found this useful?

Share it with someone who would like MonoPCC.

Comparisons