Table 3.

System1’s Test Your Ad data: Average accuracy of the video classifier broken down by input modality, number of video frames and pre-training method.

Modality	Frames	Pre-training	Accuracy
RGB	4	Imagenet	38.8
RGB+audio	4	Imagenet	42.4
RGB	8	Imagenet	38.9
RGB+audio	8	Imagenet	43.2
RGB	16	Imagenet	38.8
RGB+audio	16	Imagenet	42.2
RGB	4	INET21K	39.7
RGB+audio	4	INET21K	43.4
RGB	8	INET21K	41.2
RGB+audio	8	INET21K	43.5
RGB	16	INET21K	40.9
RGB+audio	16	INET21K	43.6