Large Displacement Optical Flow with Adaptive Feature Match

This paper presents an accurate large displacement optical flow estimation approach by adaptively integrate the local feature match. Despite coarse-to-fine warping approach can handle large displacement optical flow; however, there is inherent limit for small object with large motion. And recently integration of feature match to the variational framework has relaxed the limit, but raised another problem of ambiguous feature matching due to poor feature descriptor. Address the aforementioned problem, in this paper we propose an adaptive integration approach of local feature match. The essence is that we only keep the robust feature and remove those unstable features (e.g, textureless region) to improve the flow accuracy. The adaptive approach substantially decreases the computational cost by remove uncertain features and leads to more robust performance by excluding unreliable matches. We qualitatively and quantitatively compared to the conventional flow methods on Middlebury and Sintel benchmark and show that we achieve more accurate and promising results.