Embedded Files
For evaluate the performance of enhancement approaches against real-word corruptions, we use four enhancement approaches that contain Style Transfer, Contrast Limited Adaptive Histogram Equalization, Non-Local Means and Adversarial Traning.
The following figures show the output of those approaches.
The test result of enhancement and adversarial training on real-world corruptions is show in this table. The results are the average values of coco and bdd100k. Except for the Clean column, the numbers in other columns indicate that the change value of CE(corruption error) is compared with the clean data set. CE greater than zero means an increase, and less than zero means a decrease.
The model enhancement time are shown in the following table. The Y indicates the model have enhanced. The number in brackets means enhance model time. The unit of enhancement time is 24 hours. The dataset contain coco and bdd100k.
Page updated
Google Sites
Report abuse