Temporal objectness: Model-free learning of object proposals in video
Liang Peng, Xiaojun Qi · 2016
Intrinsic natures of different appearance between sub-regions of objects and non-objects in optical flows lead to more visual consistency for object proposals. Hence, visual variations in different sub-regions in video sequences over time is a good indicator for likeliness of objects. We propose a method that dynamically measures the objectness of each proposal by exploiting temporal consistency within each optical flow. We develop a block-based feature representation using object's spatial property and define an objectness measure using the temporal changes of this spatial representation. As a result, the proposed temporal objectness learns good object proposals over a short period (e.g., less than 1 second). The proposed method is model-free and can be used to simultaneously learn and track object proposals without training. Experiments on a video dataset shows that the proposed approach significantly outperforms state-of-the-art methods in terms of precision-recall.