🥇 CVPR 2025 Winner - SoccerNet Monocular Depth Estimation Challenge

🏟️ Project Description

This repository contains the winning solution to the SoccerNet Monocular Depth Estimation Challenge at CVPR 2025. Our method predicts accurate depth maps from RGB images of soccer matches, enabling improved 3D spatial understanding in sports analytics.

📈 Challenge Results

🏆 1st Place at the SoccerNet Monocular Depth Estimation Challenge (CVPR 2025)

Rank	Team	RMSE	AbsRel	RMSElog	SqRel	SIlog
1️⃣	Hands-On Computer Vision	0.00242	0.00164	0.00432	2e-05	0.43
2	HUST-iPad	0.00258	0.00179	0.00468	3e-05	0.47
3	bupt miclab	0.00268	0.00186	0.00484	3e-05	0.48
4	jacekm	0.00275	0.00207	0.00500	3e-05	0.50
5	hvrl	0.00282	0.00228	0.00502	3e-05	0.50

📺 Presentation Video
📊 Official Leaderboard

---

🔍 Key Features

• 🔄 Based on Depth Anything V2, a state-of-the-art model for depth estimation.
• 📊 Evaluated using standard metrics: RMSE, AbsRel, SILog, RMSElog, SqRel.
• 🧠 Fine-tuned for soccer-specific imagery and temporal consistency.
• 🖼️ Preserves sharp edges and fine details via specialized losses.
• 📏 Full-resolution training for enhanced accuracy.

🧩 Architecture

The solution uses the Depth Anything V2 architecture with different encoders (Small, Base, Large), which combines a pre-trained ViT backbone with a specialized DPT head to estimate precise metric depth.

🛠️ Requirements

• Python 3.9
• PyTorch
• CUDA (for accelerated training)

📂 Repository Structure

• data/: Data loaders and utilities
• loss/: Specialized loss functions
• metrics/: Evaluation metrics implementation
• models/: Model architectures (Depth Anything V2)

📊 Results

Config	Abs Rel ×10⁻³ ↓	RMSE ×10⁻³ ↓	RMSE Log ×10⁻³ ↓	Sq Rel ×10⁻⁴ ↓	SILog ↓
Best Baseline [Leduc et al., 2024]	2.429	2.343	4.002	0.121	0.400
Ours Half-Res	2.781	2.516	4.350	0.125	0.435
Ours Full-Res	1.443	1.590	2.619	0.062	0.261

🚀 Usage

# Training
bash train.sh

# Evaluation
python evaluate_depth.py --pred_dir [predictions_directory] --gt_dir [ground_truth_directory]

# Inference
python challenge.py --encoder vitl --checkpoint_path [checkpoint_path] --input_dir [input_directory] --output_dir [output_directory]

Training options can be modified in the train.sh file.