Blur generation and object detection benchmark

[KengChiLiu]

Hi @KupynOrest , thanks for your work. Another two questions here (1) Do you think it is OK for you to provide the ground truth labels of your detection benchmark ??? (2) Are there any source codes for blurring generation of your detection benchmark ? In your paper (6.3. Object Detection benchmark on YOLO) it says that "Following [13][23][25] we take a random between 5 and 25 frames taken by 240fps bla bla bla". It seems to be different from your (4. Motion blur generation) and the code you have provided.

Thank you~

[KupynOrest]

Hello, the ground truth labels were just the predictions os YOLO network on sharp images, so that the idea would be to get as close predictions to them from deblurred photos, we have published the larger dataset online, and for the evaluation we manually selected ~400 frames from it with clear cars on the photo.

• We will update the code this or next week, and provide the code for blur generation, for this dataset we used standard approach to simulate blur from high frame rate camera, in contrast to creating synth blur as described in (4. Motion blur generation)