Fig. 10
From: Large scale analysis of dataset and simulation biases in SLAM research
Sim2Real gap: we applied different levels of blur (low, medium, high) on sequences acquired through photo-realistic simulator GTA-V. Tracking performance of state-of-the-art visual SLAM drastically degrades even with low blur. Deep deblur71 does not necessarily improve tracking performance. In some cases (medium blur) ATE degrades after deblurring. As compared to changing conditions (season, weather, times of the day, dynamic objects), where tracking was successful during entire track length (Table 7), motion blur considerably degrades visual SLAM tracking. Photo-realism in simulators does not guarantee motion blur. Therefore, these photo-realistic platforms overestimate visual SLAM performance.
