LightDepth - Lightweight Depth Estimation

LightDepth is a lightweight monocular depth estimation model, built with a ResNet18 encoder-decoder architecture, it demonstrates core concepts of depth estimation from single RGB images without the complexity of production-scale models. The project provides a clear, straightforward implementation ideal for learning about feature extraction, multi-scale processing, and depth prediction in computer vision. Compared project performance with Depth Anything V2 on the NYU Depth V2 test dataset, demonstrating relatively competitive results with significantly fewer parameters and faster inference times.

Features

• Simple Architecture: ResNet18 encoder with U-Net style decoder. Uses pretrained ImageNet weights.
• Easy Training: Straightforward training script with minimal configuration.
• L1 Loss: Simple and effective loss function.
• Multiple Metrics: RMSE, MAE, AbsRel, and SqRel metrics.
• Fast Inference: 72% faster than Depth Anything V2.
• Compact Model: 42% fewer parameters than Depth Anything V2 small (14.3M vs 24.8M).

Dataset

Download the NYU Depth v2 dataset from Kaggle.

Place the dataset in data/nyu/ directory with the following structure:

data/nyu/
├── nyu2_train.csv
├── nyu2_test.csv
├── nyu2_train/
│   ├── *.jpg
│   └── *.png
└── nyu2_test/
    ├── *.jpg
    └── *.png

Pretrained Model

Download the trained model checkpoint from Google Drive.

Training Infrastructure:

• GPU: RTX 5080 16GB
• CPU: AMD Ryzen 7 9800X3D
• RAM: 32GB
• OS: Windows 11
• Epochs: 50
• Dataset: NYU Depth v2

Project Report

For detailed methodology, experiments, and analysis, see: Project Report

Model Architecture

• Encoder: ResNet18 (pretrained on ImageNet)
• Decoder: 4-stage upsampling with skip connections
• Channels: [64, 64, 128, 256, 512] → [256, 128, 64, 32] → 1
• Loss: L1 (Mean Absolute Error)
• Metrics: RMSE, MAE, AbsRel, SqRel

Results

Quantitative Comparison

LightDepth vs Depth Anything V2 on NYU Depth v2 test set:

Metric	LightDepth	Depth Anything V2	Winner
Parameters	14,330,369	24,785,089	LightDepth (42% fewer)
RMSE ↓	2.9477	2.8074	Depth Anything V2 (5.0%)
MAE ↓	2.6758	2.2360	Depth Anything V2 (19.7%)
Absolute Relative Error ↓	0.9724	1.0272	LightDepth (5.3%)
Squared Relative Error ↓	2.6063	3.7849	LightDepth (31.1%)
Total Inference Time ↓	5 seconds	18 seconds	LightDepth (72.2% faster)

Key Findings:

• LightDepth achieves competitive performance with 42% fewer parameters
• Significantly faster inference time (72% improvement)
• Better performance on relative error metrics
• Slight trade-off in absolute error metrics (RMSE, MAE)

Qualitative Results

Sample depth predictions on NYU Depth V2 test set:

Sample 1

Input	Ground Truth	LightDepth	Depth Anything V2

Sample 2

Input	Ground Truth	LightDepth	Depth Anything V2

Sample 3

Input	Ground Truth	LightDepth	Depth Anything V2

Usage

Requirements

• Python 3.14
• PyTorch
• CUDA-capable GPU (recommended)

Installation

1. Clone the repository:

git clone https://github.com/suxrobGM/lightdepth.git
cd lightdepth

2. Install dependencies either via PDM or pip:

# Using PDM (Recommended)
pdm install

# Or using pip
pip install -r requirements.txt

Note

You can install pdm via pip install pdm if you don't have it already.

Training

You can train the model from scratch or use a pretrained model from the link above.

# Using PDM
pdm train

# Or directly with Python
python scripts/train.py --config config.yaml

Options:

• --config: Path to configuration file (default: config.yaml)
• --resume: Resume training from checkpoint file (optional)

The model saves the best checkpoint to checkpoints/best_model.pth.

Evaluation

Evaluate a trained model on the test set:

# Using PDM
pdm eval --checkpoint checkpoints/best_model.pth

# Or directly with Python
python scripts/eval.py --checkpoint checkpoints/best_model.pth --config config.yaml

Note

If you downloaded the pretrained model, use its checkpoint path. For example: --checkpoint checkpoints/lightdepth.pth

Options:

• --checkpoint: Path to model checkpoint (required)
• --config: Path to configuration file

Inference

Run inference on a single image:

# Using PDM
pdm infer --checkpoint checkpoints/best_model.pth --input image.png --output depth.png

# Or directly with Python
python scripts/infer.py --checkpoint checkpoints/best_model.pth --input image.png --output depth.png

Note

If you downloaded the pretrained model, use its checkpoint path. For example: --checkpoint checkpoints/lightdepth.pth

Options:

• --checkpoint: Path to model checkpoint (required)
• --input: Path to input RGB image (required)
• --output: Path to save depth map (default: output/depth.png)
• --colormap: Colormap for visualization (plasma, viridis, magma, inferno, gray). Use gray for grayscale output.

Comparison with Depth Anything V2

Compare LightDepth with Depth Anything V2:

pdm compare --checkpoint checkpoints/best_model.pth

Note

If you downloaded the pretrained model, use its checkpoint path. For example: --checkpoint checkpoints/lightdepth.pth

Options:

• --config: Path to configuration file
• lightdepth-checkpoint: Path to LightDepth model checkpoint (required)
• --dav2-model: Depth Anything V2 model. Applicable options: depth-anything/Depth-Anything-V2-Small-hf, depth-anything/Depth-Anything-V2-Base, depth-anything/Depth-Anything-V2-Large (default: depth-anything/Depth-Anything-V2-Small-hf)
• --output: Path to save comparison results JSON (default: output/comparison_results.json)
• --visualize: Number of samples to visualize (0 to disable)

Configuration

Edit config.yaml to change training settings. It has well-documented parameters for easy customization.

Project Structure

lightdepth/
├── src/lightdepth/
│   ├── models/          # Model architectures
│   │   ├── encoder.py   # ResNet18 encoder
│   │   ├── decoder.py   # U-Net decoder
│   │   └── lightdepth.py # Complete model
│   ├── data/            # Data loading
│   │   ├── dataset.py   # NYU dataset
│   │   ├── dataloader.py # DataLoader
│   │   └── transforms.py # Data augmentations
│   └── utils/           # Utilities
│       ├── config.py    # Configuration
│       ├── losses.py    # L1 loss
│       └── metrics.py   # RMSE metric
│       └── visualization.py # Visualization utilities
├── scripts/
│   ├── train.py         # Training script
│   ├── eval.py          # Evaluation script
│   ├── infer.py         # Inference script
│   └── compare_models.py       # Comparison script with Depth Anything V2
├── config.yaml          # Default training and inference configuration
├── requirements.txt     # Python dependencies for pip
└── pyproject.toml       # Python dependencies for PDM

License

This project is licensed under the MIT License - see the LICENSE file for details.

Author

• Sukhrobbek Ilyosbekov
• CS 7180 Advanced Perception

Citation

If you use LightDepth in your research or project, please cite:

@misc{lightdepth2025,
  title={LightDepth: Lightweight Depth Estimation},
  author={Sukhrobbek Ilyosbekov},
  year={2025},
  url={https://github.com/suxrobGM/lightdepth}
}

This project was inspired by and compared against Depth Anything V2:

@article{depth_anything_v2,
  title={Depth Anything V2},
  author={Yang, Lihe and Kang, Bingyi and Huang, Zilong and Zhao, Zhen and Xu, Xiaogang and Feng, Jiashi and Zhao, Hengshuang},
  journal={arXiv:2406.09414},
  year={2024},
  url={https://arxiv.org/abs/2406.09414}
}