Evaluation of Camera Resolution in Optical Flow Estimation Using Event-Based Cameras

Abstract: Developments in event-based camera technology and their recent increase in pixel count raised the question of whether resolution helps the accuracy and performance of algorithms. This thesis studies the impact of resolution on optical flow estimation for event-based cameras. For this purpose, we created a data set containing a mix of synthetic scenes and real camera recordings with ground truth available. For the modeling of low-resolution data, we designed three different downsampling algorithms. The camera used for the real scene recordings was the Prophesee (CSD3SVCD), which was determined to be the best out of the current state-of-the-art cameras in a prestudy. The camera investigation evaluated the camera’s performance in terms of temporal and spatial accuracy. In order to answer the question, whether resolution benefits the accuracy of optical flow estimation, we ran a total of 13 algorithms variations from four algorithm families (Lucas-Kanade [1, 2], Local-Planes fitting [2, 3], direction-selective filter [2, 4] and patch match [5]) on the data set. We then analysed their performance in terms of processing time, output density, angular error, endpoint error and relative endpoint error. The results show that no global correlation between resolution and accuracy across all algorithms can be identified. However, methods show individually different behaviour on different data. The best performing methods, the patch match algorithms, seemed to prefer the less dense downsampled data. The evaluation also showed that rather than resolution, the specific characteristics of the data seemed to have a larger impact on accuracy. Thus denoised data might increase accuracy more than a change of resolution.