Final-Projects-Depth-Aware Bokeh

This project implements a depth bokeh system and web demo. Given a single RGB image, we estimate a depth map and apply a depth-based blur to simulate camera aperture effects. The project also includes a small web demo and an evaluation script to compare results against different baselines.

⚙️ Installation

# if you would like to use GPU
conda create -n depthbokeh python=3.10 -y
conda activate depthbokeh
pip install torch==2.3.0+cu121 torchvision==0.18.0+cu121 --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt

# if you are not using GPU
pip install -r requirements.txt

Tree

.
├── dataset/
│   ├── train/
│   │   ├── gt/            # unblut image
│   │   └── depth/         # generated depth maps (.npy)
│   └── val/
│       ├── baseline/      # ground truth images for evaluation
|       ├── depth/         # generated depth maps (.npy)
│       ├── result/        # our method outputs (40r.jpg)
│       ├── google/        # google baseline outputs (40g.jpg)
|       ├── gt/            # original RGB images without blur
│       └── gpt/           # gpt baseline outputs (40gpt.jpg)
│
├── model/
│   ├── blur.py            # depth  bokeh rendering
│   └── depth_generate.py  # depth estimation (DepthPro / Depth-Anything)
│
├── src/
│   ├── utils.py           # all helper
│   ├── gpt_gen.py         # generate gpt images
│   └── evaluate.py        # PSNR / SSIM / LPIPS evaluation
│
├── templates/
│   └── index.html         # web to generate our result
│
├── server.py              # Flask web demo
├── requirements.txt
├── evaluation.py
├── run_interactive.py
└── README.md

Depth Map Generation

We support two pre-trained depth models:

• High quality: Apple DepthPro
• Low quality but faster: Depth-Anything V2

Depth maps will be saved as .npy files.

If you would like to generate a depth map you can:

# change the path!
python model/depth_generate.py

Web Demo (Depth-Aware Bokeh)

We have a small Flask server provides an interactive demo.

The features are:

1. upload an image
2. click to select focus point
3. adjust aperture to control blur strength

Run the server:

python app.py

Open in your browser:

http://localhost:5000

Evaluation (PSNR / SSIM / LPIPS)

We evaluate results using:

• PSNR – pixel-level reconstruction accuracy
• SSIM – structural similarity
• LPIPS – perceptual similarity

The evaluation compares:

• our method (using our demo code)
• a Google phone editing
• a GPT-based generative baseline (Gpt5)

Expected file naming:

• GT: baseline/40.jpg
• ours: result/40r.jpg
• google: google/40g.jpg
• gpt: gpt/40gpt.jpg

Run evaluation

Open src/evaluate.py set your base path Then run:

python src/evaluate.py

Average PSNR / SSIM / LPIPS values are printed to the terminal.