We include two videos in our supplementary material showing results from EDnCNN on the three datasets tested in our paper. Both DVS OF and IROS18 were collected from a different camera model than DVSNOISE20 (training source), but EDnCNN still performed well showing that the methodology is robust to cameras models. 

1. DVSNOISE20 Dataset (Ours) (DVSNOISE20.mp4) 
All datasets show good qualitative results for denoising - excluding "Pavers". As shown in Figure 5, the "Pavers" scene proved challenging for all denoising methods. This is likely due to the high amount of clutter and minimal scene contrast. These two factors limit spatial and temporal consistency that most algorithms rely upon for denoising. As stated in the paper, EDnCNN was trained using a leave-one-scene-out method to ensure the network was not trained on specific scene content.

2. DVS Optical Flow (OF) Dataset (OtherDatasets.mp4)
This video shows two short scenes of unrestricted camera motion. EDnCNN was trained using only DVSNOISE20 data, but still performed very well at denoising. The only input from this dataset to EDnCNN was the DVS events (i.e. APS frames and IMU information are ignored).

3. IROS18 Dataset (OtherDatasets.mp4)
This dataset, designed for moving object detection and tracking, consists of seven different scenes and included both unrestricted camera and object motion. The only input from this dataset to EDnCNN was the DVS events (i.e. APS frames and IMU information are ignored).

