| dc.description.abstract | Object detection from images captured by Unmanned Aerial Vehicles (UAVs) is becoming dramatically useful. Despite the great success of the generic object detection methods trained on ground-to-ground images, a huge performance drop is observed when these methods are directly applied to images captured by UAVs. The unsatisfactory performance is owing to many UAV-specific nuisances, such as varying flying altitudes, adverse weather conditions, dynamically changing viewing angles, etc., constituting a large number of fine-grained domains across which the detection model has to stay robust. Fortunately, UAVs record meta-data corresponding to the same varying attributes, which can either be freely available along with the UAV images, or easily obtained. We propose to utilize the free meta-data in conjunction with the associated UAV images to learn domain-robust features via an adversarial training framework. This model is dubbed Nuisance Disentangled Feature Transforms (NDFT), for the specific challenging problem of object detection in UAV images. It achieves a substantial gain in robustness to these nuisances. This work demonstrates the effectiveness of our proposed algorithm by showing both quantitative improvements on two existing UAV-based object detection benchmarks, as well as qualitative improvements on self-collected UAV imagery.
Reprinted with permission from the Abstract section of Deep Nuisance Disentanglement for Robust Object Detection from Unmanned Aerial Vehicles by Zhenyu Wu† , Karthik Suresh† , Priya Narayanan, Hongyu Xu, Heesung Kwon, Zhangyang Wang, 2019, International Conference on Computer Vision (ICCV 2019) Proceedings (Under Review). † indicates equal contribution | en |
