AdaptSyS Dataset

The AdaptSyS dataset demonstrates use of the Adaptive Synthset Method detailed in the "Adaptive Synthset System for Person Detection and Raycast-Based Geolocation in UAV-Based SAR Missions" paper. The AdaptSyS dataset consists of 4 synthsets, generated with 3 different generation methods, with and without shadows. Each synthset contains 3,000 RGB images with annotations (in YOLO format), split into train (2,400 images) and validation (600 images) subsets: Synthset Generation method Ground shadow AS-DCC-v4 DCC YES AS-GAME-v1 game engine YES AS-GAME-v2 game engine NO AS-COMP-v2 compositing NO All synthset use random colors for 3D person texturing, 9 different animation states (a combination of mo-cap and ragdoll-based crowd simulation), and background tiles from the HERIDAL-384 dataset. Additionally, the AS-COMP-v2 synthset includes, within the _pre-comp folder, annotated raw renderings (without ground shadows) of 3D persons with an alpha channel, ready for compositing over any background.