Computing HDR images from a sequence of 100 noisy images captured by a 14-bit handheld camera. The noise in the input images is higher in dark regions, as shown in the tone-mapped images. Our approach produces sharp and clean HDR images, and works for complex scenes with large depth variation.

One of 100 noisy input images, PSNR=35.8132 dB
Tonemapped noisy input image
Tonemapped result of robust averaging along flow (followed by CBM3D), PSNR=43.281 dB
Tonemapped result of PCA denoising along flow (followed by CBM3D), PSNR=44.9549 dB
Tonemapped Ground Truth

The trajectories of two sample feature points in the scene calculated from optical flow computation. The noticeable difference in the trajectories clearly indicates spatially-varying motion, which is difficult for deblurring approach to work effectively on this scene.

Two sample points marked in red and green
Motion paths corresponding to the red and green points in the scene