Title: Progressive Domain Adaptation with Contrastive Learning for Object Classification in the Satellite Imagery
Student: Biswas, Debojyoti firstname.lastname@example.org
Advisor: Jelena Tešić
Date and Time: Thu Dec 8 2022 2:00 PM
Abstract: Recent increases in aerial imagery capture rates and increase in the computational power of on-board processing have opened the door to scaling up object detection and domain adaptation research to production. The aerial imagery domain is a difficult domain to transition to from well-established consumer imagery analysis production for two reasons: (a) data sets are enormous in size, and each frame contains many dense and small objects; and (b) high variation between data sets due to altitude, weather, rapid change of scenery, object classes, and object sizes. Thus, object detectors trained on one set of labeled aerial imagery data are likely to perform poorly for unseen dataset analytics. In this study, we present the progressive domain adaptation (DA) method that uses contrastive learning to overcome large variations in aerial data imagery. We show that feature alignment is an effective way for domain adaption, and we introduce feature alignment in local and global feature space from the Feature Pyramid Network (FPN) output. Next, we use contrastive learning to place similar images close to each other and dissimilar images distant in the feature space across domains. We demonstrate improved mAP for difficult classes on domain adaptation of models learned in DIOR to DOTA2.0 dataset.
Deadline: Dec. 23, 2022, midnight