HDR Denoising and Deblurring by Learning Spatio-temporal Distortion Models

1. Additional Results for HDR Reconstruction

This section shows additional HDR results captured with various exposure times. Different networks are used for particular exposure rates that we consider: 1:16 (statue), 1:8 (snow), and 1:4 (all remaining scenes). However, same network is used for different absolute exposure times having the same rate. The blur in the scenes is caused by either motion of the objects or camera movement.

Candle Car Ceiling Clouds Laptop Window Cards Statue Snow


2. Temporal super-resolution

This section illustrates temporally created 4 sharp HDR frames using our temporal super-resolution network.

Example Secene 1:

Frame 1 Frame 2 Frame 3 Frame 4

Example Secene 2:

Frame 1 Frame 2 Frame 3 Frame 4

3. Blur Adjustment

This section explains continous adjustment of motion blur for the single HDR frame. There is a continous blur information in high exposure that we cut it into individual sharp frames which emulate certain window of time. Then, blur is achieved by averaging temporally reconstructed those sharp HDR frames. The amount of blur depends on the number of averaged frames which means addition of all frames simulate the exact blur information in high exposure. As a result, objects are blurred naturally along the motion direction in our reconstructions. The exposure rates of the scenes are as follows: 1:8 (rosary, clouds) and 1:4 (all remaining scenes).

Cup Pen Marble Candle Rosary Clouds

Blur setting:

Low/High Exposure Time: 8ms/32ms

4. Video

This section demonstrates reconstructed HDR videos captured in dynamic scenes. Videos are captured with exposure rates of 1:16 and 1:4.

Video 1 Video 2

Low/High Exposure Time: 1ms/16ms
Imprint / Data Protection