Scene shape estimation from multiple partly cloudy days

Cloudmaps from static ground-view video

“…When an accurate 3D scene model is available, the known-geometry method is most appropriate as it will be more accurate than the planar method and is simpler than the direct method. Such geometry can be estimated automatically from video data using a variety of techniques, including cloud shadow motion [20,21], but these methods require more video (often from multiple days) to obtain a high-quality scene model. If the camera will be used repeatedly to estimate a cloudmap, it may be beneficial to obtain a 3D scene model using other methods, such as structure from motion or laser scanning, as errors in the scene model can introduce artifacts in the resulting cloudmap.…”

“…Recent work has shown the benefits of explicitly reasoning about the underlying causes of outdoor appearance variations. For example, color changes due to sun motion are strong cues to scene shape and albedo [4,19,5,6] and transient clouds and their shadows are strong cues to scene geometry [20,21] and camera calibration [22]. Outdoor videos have also been used to estimate dynamic scene properties, including low-dimensional cloudiness models [7,8,9].…”

“…There are many approaches [30,31] for estimating the temporal delay, δ pq , between two time series, I(p, t) and I(q, t). We compute the delay using a two-stage process, first described by Workman et al [21]. To initialize, we select the integer delay offset that maximizes the correlation between the two time series (in our implementation we use the time series of the green color channel, because it had less spatial noise than the intensity in the videos we use for evaluation).…”