1st Place 🏆 Solution Report of team CV for the Challenge "Core Values: A Geological Quest"

Challenge Website: Link

Team Member: Efstathios Karypidis

Contact: stathiskaripidis@gmail.com

Short Description

Our solution implements a two-phase methodology for geological image analysis. In the first phase, I developed a specialized neural network architecture leveraging state-of-the-art Vision Foundation Models, trained on both labeled and unlabeled datasets. The second phase employs an advanced prediction refinement pipeline that integrates the Segment Anything Model 2 (SAM2) to enhance segmentation accuracy and robustness.

Installation

1. Create a new conda environment and install the required packages:

conda create -n cvgq python=3.11
conda activate cvgq
pip3 install torch torchvision torchaudio

2. Clone and install Segment Anything Model 2 (SAM2) from the official repository:

git clone https://github.com/facebookresearch/sam2
cd sam2
pip install -e .

3. Sam has a problem and configs should be placed in the parent sam2 folder. Inside the parent sam2 folder, run the following command:

cp -r sam2/configs/ .

4. Download sam2 checkpoints

cd checkpoints
bash download_ckpts.sh
cd ../.. # Return to the root folder

5. Download the pretrained weights from the release page. Select the file pretrained_models.zip and place the checkpoints in the checkpoints and final_checkpoints folders.

Structure

The repository is structured as follows:

cvs
│
├───data
│   ├───core-values-test-data
│   ├───train
│   ├───train_unlabeled
├───checkpoints
├───src
│   ├───model.py
│   ├───data.py
│   ├───transforms.py
│   ├───utils.py
│   ├───my_utils.py
├───pred_vis
├───train.py
├───train_semi.py
|───predict.py 
├───solution.ipynb
|───README.md

• data: contains the dataset folders. Download the dataset from the official challenge website and unzip the files. Note that file image001269.png from the unlabeled dataset is corrupted and is removed.
• checkpoints: folder to save the trained model checkpoints.
• final_checkpoints: contains the the 5 trained model checkpoints used for the final submission.
• src: contains the source code files.
• model.py: contains the neural network architecture
• train.py: script to train the model on the labeled dataset.
• train_semi.py: script to train the model on both labeled and unlabeled datasets.
• utils.py: contains utility functions provided by the challenge organizers github repository.
• my_utils.py: contains utility functions implemented by me.
• predict.py: script to make predictions on the test dataset.
• solution.ipynb: Jupyter notebook containing the solution report.
• pred_vis: folder to save the prediction visualizations (I run the prediction script on the public test dataset and saved the predictions).
• README.md: this file.

Usage

To understand and reproduce my solution, please refer to the solution.ipynb notebook. The notebook provides a detailed explanation of the methodology, implementation, some insights and results.

Incremental Improvements

In order to show important incremental improvements, I provide a table with the experiments conducted during the development of the solution. The table includes the experiment id, description, public dice score, and the checkpoint file.

Experiment Id	Description	Public Dice Score	Checkpoint
1	Supervised Only (10 epochs, CE loss)	0.5569	model_supervised.pth
2	Semi-Supervised (25 epochs, CE loss)	0.5910	---
3	Semi-Supervised (50 epochs, CE loss)	0.6058	---
4	Semi-Supervised (100 epochs, CE loss)	0.6413	---
5	Semi-Supervised (150 epochs, CE loss)	0.6422	model_semi_blr2em5.pth
6	Semi-Supervised (150 epochs, blr 4e-5, CE loss)	0.6441	model_semi.pth
7	Semi-Supervised (100 epochs, blr 4e-5, Ohem loss)	0.6445	model_semi_100eps_ohem.pth
8	Semi-Supervised (150 epochs, blr 4e-5, Consistency Reg)	0.6473	model_semi_creg.pth
9	Semi-Supervised (150 epochs, blr 4e-5, Combined loss)	0.6486	model_semi_combloss.pth
10	Ensemble of 6, 8, 9	0.6599	---
11	Ensemble of 6, 7, 8, 9	0.6642	---
12	Ensemble of 5, 6, 7, 8, 9	0.6655	---
13	Ensemble of 5, 6, 7, 8, 9, 11 + SAM2 AutoMaskGen Refine	0.6718	---
14	Ensemble of 5, 6, 7, 8, 9, 11 + SAM2 AutoMaskGen Refine + Replace Confident(thresh 0.8)	0.6868	---
15	Ensemble of 5, 6, 7, 8, 9, 11 + SAM2 AutoMaskGen Refine + Replace Confident(thresh 0.625)	0.6951	---

Candidate for Honorable Mention

• Fastest GPU Inference. My algorithm performs inference on 70 images of the public test dataset in ~ 4 minutes.
• Fastest CPU Inference. Not tested.
• Best custom model that does not use a highly abstracted API. The main model architecture is implemented in PyTorch. Only for prediction refinement, I use the SAM2AutomaticMaskGenerator.
• Most innovative use of unlabelled data. I use state-of-the-art semi supervised learning techniques to leverage the unlabeled data.
• Best documentation. The solution report is detailed and provides a clear explanation of the methodology, implementation, and results.