We present RRODT, a real-world rainy video dataset for testing the efficacy of deraining methods to help downstream object detection and object tracking. RRODT is composed of 57 videos and totally 10258 frames. It consists of two kinds of annotations. One set of annotation contains 33077 objects for detection. The other set contains 408 unique objects for tracking.
We create a RainVID&SS benchmark on which the performance gain of subsequent processes with and without deraining pre-process can be evaluated. We argue that the gain can be a metric of deraining methods from a computer perspective