Synthetic Defocusing using Monocular Depth Estimation
Graded blurring of an image based on how far a point is from focus using just a single image with no extra data. To achieve this we generate a depth map for the image using machine learning. Create blurred versions of the image in levels based on how far a point is from the depth of focus. Stitch the different blurred images to create the final image with the selected point fully focussed and points further away become more blurred.
You'll need to setup Tensorflow v1 and OpenCV for the bare minimum run.
NOTE: Setting up Tensorflow v1 can be tricky, hence it is recommeneded to use a prebuilt docker image.
docker pull tensorflow/tensorflow:1.0.0-py3GPU specific images are also available.
Install a trained model for depth estimation.
sh ./depth/get_model.sh model_kitti depth/This will download the model which will give the best results for the sample in the repo.
Call main.py with model path and image path
python main.py --model_path /path/to/model --image_path /path/to/imageThe image will load up. Clicking anywhere sets that as the point of focus, regenerating the image.
I've actually pushed in the depth data for the sample images, so you can actually run the program without tensorflow or a trained model for the sample images :)
Install dependencies from requirements.txt
Run the gui version as
python gui.py --model_path /path/to/modelGui version requires some additional packages. Install as necessary.
We have used the method based on the paper Unsupervised Monocular Depth Estimation with Left-Right Consistency. You can find more about their amazing paper here. They train a machine learning model to generate a depth map using just a single image.
The model is trained on a large set of stereo(left-right) images to generate a right image from a given left image. Generating the right image, the model is learning internally about the depth of various points in the image.
Once the model is trained. We can give it a simple image and it will be able to generate a depth map for that image.
You can learn more about their amazing project on github. Their trained models give way better results than what we can so we are using that. The depth estimation code we have is a minimal stripped down version just to run the model.
This was done as a final year project by Haritha Paul, Navin Mohan, Roshan V and me.
Leave a star if you liked the project. :)