AIM 2020 Challenge on Rendering Realistic Bokeh

Core idea

The AIM 2020 Bokeh Effect Rendering challenge tasked participants with synthesizing realistic shallow-depth-of-field ("bokeh") images from regular all-in-focus photos — a hard image-to-image translation problem requiring accurate depth-aware blur.

Approach

The team proposed a modified U-Net architecture for the task. Compared to the original U-Net, the max-pooling downsampling operation was replaced with a strided convolution layer, and the feature maps from the shortcut (skip) connections are concatenated before applying the activation function in the decoder module. Leaky ReLU activations are used throughout the convolutional layers, and the entire model is trained end-to-end to minimize mean absolute error (MAE) loss using the Adam optimizer.

My contribution

I was part of a 2-member team behind this submission during my undergraduate research work in the Computer Vision Lab under Dr. Jiji C V. Our entry was evaluated against other international submissions on the official challenge benchmark and ranked 6th.

Published as a chapter in Computer Vision – ECCV 2020 Workshops (Springer, Cham), pp. 213–228. DOI: 10.1007/978-3-030-67070-2_13