Unsupervised Action Localization Crop in Video Retargeting for 3D ConvNets

Untrimmed videos on social media or those captured by robots and surveillance cameras are of varied aspect ratios. However, 3D CNNs require a square-shaped video whose spatial dimension is smaller than the original one. Random or center-cropping techniques in use may leave out the video's subject altogether. To address this, we propose an unsupervised video cropping approach by shaping this as a retargeting and video-to-video synthesis problem. The synthesized video maintains 1:1 aspect ratio, smaller in size and is targeted at the video-subject throughout the whole duration. First, action localization on the individual frames is performed by identifying patches with homogeneous motion patterns and a single salient patch is pin-pointed. To avoid viewpoint jitters and flickering artifacts, any inter-frame scale or position changes among the patches is performed gradually over time. This issue is addressed with a poly-Bezier fitting in 3D space that passes through some chosen pivot timestamps and its shape is influenced by in-between control timestamps. To corroborate the effectiveness of the proposed method, we evaluate the video classification task by comparing our dynamic cropping with static random on three benchmark datasets: UCF-101, HMDB-51 and ActivityNet v1.3. The clip accuracy and top-1 accuracy for video classification after our cropping, outperform 3D CNN performances for same-sized inputs with random crop; sometimes even surpassing larger random crop sizes.