Computer Vision News - February 2019

Motion detection: Motion detection (Optical Flow) is a challenging problem because the camera is in (fluctuating) motion simultaneously with the movement of objects the camera captures. With the detection of moving objects being of the utmost importance in Automated Driving to avoid accident casualties and fatalities of autonomous vehicles. This process uses motion cues to enable detection of a generic object, because (for the time being) there are no network to detect all the possible objects. Classic motion detection approaches focused on geometry-based methods, purely geometry-based approaches, however, suffer from many limitations. One such limitation is the problem of parallax -- the change in the perceived location of two points in relation to each-other, from the viewer’s point of view, caused by a change in the viewer’s location. Leading work in this field includes: Fragkiadaki et al’s proposal for segmenting moving objects using a separate proposal generation approach; however, this is computationally inefficient. Jain et al’s motion fusion based method focused on a generic object. Tokmakov et al’s use of a fully convolutional network with optical flow to estimate type of motion, which can work with either optical flow or concatenated image; note that concatenated image inputs won’t benefit from the pre-trained weights, as those were acquired using RGB inputs only. Finally, Drayer et al’s video segmentation work applied R-CNN methods followed by object segmentation using a spatio-temporal graph. Illustration of dense Motion detection (Optical Flow) CNN Based Pipelines: The nature of Deep Learning networks allows using a shared architecture to jointly learn different regression tasks - this is especially useful for critical real- time applications like Automated Driving. Here, too, there are supervised and unsupervised approaches. Tateno et al proposed a supervised approach with a Research 8 Research Computer Vision News