ECCV 2020 Daily - Wednesday

3 Razieh Kaviani 17 finer-scale representations to recognize unknown pixels better.” Razieh discovered another challenge for satellite images is that there are some similar spectrals. Therefore, before projecting the data into the abundance space, they add another step to project the data from the raw image space through the embedding feature space with small discriminating features. To find outmore about Razieh and Ying’s work, check out their presentation video, andvisit their poster presentation [#4443] tomorrow (Thursday) at 14:00 and Friday at 00:00 (UTC +1). There are already some open-set recognitionmethods for natural images. However, these methods do not work well here. A crucial issue at play is ‘mixed pixel’. Each pixel in the satellite image covers a large area – maybe 30 square meters – and may be a mix of materials, such as trees, grass, or soil. This can be very confusing for the state- of-the-art methods, which work well on the raw image domain, but cannot handle satellite images. Ying tells us how they address this challenge: “We project the image on to a new domain called the abundance domain . These mixed pixels are assumed to be a linear combination of a few spectral bases with the corresponding mixing coefficients – meaning how many proportions of each spectral basis contribute to the mixtures. We found that when we study the base coefficients on the abundance space, that provides DAILY W e d n e s d a y Ying Qu