Learning Virtual Exemplars for Label-Efficient Satellite Image Change Detection

Abstract

In this paper, we devise a novel interactive satellite image change detection algorithm based on active learning. The proposed framework is iterative and relies on a question & answer model which asks the oracle (user) questions about the most informative display (subset of critical images), and according to the user's responses, updates change detections. The contribution of our framework resides in a novel display model which selects the most representative and diverse virtual exemplars that adversely challenge the learned change detection functions, thereby leading to highly discriminating functions in the subsequent iterations of active learning. Experiments, conducted on the challenging task of interactive satellite image change detection, show the superiority of the proposed virtual display model against the related work.

Deep-Net Inversion for Frugal Satellite Image Change Detection

“…on D t+1 ∪ • • • ∪ D 0 where D t+1 is a subset of critical virtual exemplars (denoted as {D k } K k=1 ) whose labeling is frugally obtained from the oracle. In contrast to our previous work [26], we consider a variant which defines for each virtual exemplar D k a distribution {µ ik } n i=1 that measures the conditional probability of associating D k to the n-training samples. With this variant, the virtual exemplars together with their distributions {µ ik } i,k are obtained by minimizing the following surrogate problem…”

“…here Ω c = {µ : µ ≥ 0; µ 1 n = 1 K } guarantees the column-stochasticity of the memberships {µ ik } i instead of row-stochasticity in [26], and 1 K , 1 n denote two vectors of K and n ones respectively. In Eq.…”

“…the unlabeled samples in I. This term, as defined, has a major advantage compared to [26]; indeed, due to column-stochasticity of µ, the first term in Eq. 1 reaches (more easily) smaller values than its counterpart in [26] (whose µ is row-stochastic), and this allows producing virtual exemplars closer to the input data and thereby inheriting more accurate labels when the oracle annotates their closest input data.…”

“…The second term in Eq. 1 also captures diversity, not in the virtual exemplars themselves (as in [26]), but instead in how training data polarize (or attract) the virtual exemplars; this term reaches its minimum when all the training samples evenly attract the virtual data. The third and fourth terms are similar to [26].…”

“…1 also captures diversity, not in the virtual exemplars themselves (as in [26]), but instead in how training data polarize (or attract) the virtual exemplars; this term reaches its minimum when all the training samples evenly attract the virtual data. The third and fourth terms are similar to [26]. The former measures the ambiguity (or uncertainty) in D as the entropy of the scoring function (it reaches its smallest value when virtual exemplars in D are evenly scored w.r.t different classes) while the latter acts as a regularizer which favors uniform conditional probability distribution {µ ik } i .…”