CDrone

The CDrone project focuses on Monocular 3D Object Detection from various camera perspectives, including car, traffic, and specifically, drone views. It addresses the limitation of existing techniques that perform well only on a limited set of benchmarks. The project's main goal is to encourage a more extended evaluation of 3D detection frameworks across different camera angles.

The project makes two key contributions. First, it introduces the CARLA Drone dataset (CDrone), a synthetic dataset simulating drone views that expands the diversity of camera perspectives in existing benchmarks. This dataset presents a real-world challenge, as previous techniques struggle to perform well on both CDrone and real-world drone datasets.

Second, CDrone develops GroundMix, an effective data augmentation pipeline. GroundMix utilizes the ground for creating 3D-consistent augmentation of training images, significantly boosting the detection accuracy of lightweight one-stage detectors. The research demonstrates on-par or substantially higher performance than previous state-of-the-art methods across all tested datasets.