無人機大量物件偵測與計算以及公開資料集
Along with the acceptance of ICCV 2017, we are happy to announce the first drone-based object (e.g., vehicle) counting dataset available for public use. The Car Parking Lot Dataset (CARPK) contains nearly 90,000 cars and annotations from 4 different parking lots collected by drones (PHANTOM 3 PROFESSIONAL) at approximate 40 meters height.
For leveraging drones for productive object localization and counting, we address the first few object counting problem by leveraging spatial layout information (e.g., cars often park regularly) and introducing the spatially regularized constraints into the neural networks for improving the localization accuracy. We show the proposed method outperforming strong baselines such as YOLO, Faster RCNN, etc., in terms of localization recall and counting accuracy.
We acknowledge the multi-year project supports from MediaTek and the DGX-1 supercomputer and grants from NVIDIA.
為了善用無人機在交通、管理、安全上的應用，我們首次成功設計無人機視角的（大量）物件計算以及偵測演算法，並利用了重複物件之間的位置相關性來設計新的卷積網路。為了驗證這個嶄新的應用，我們標註了將近九萬個的車輛物件，為目前唯一支持相關研究的資料集，並且公開讓研究社群使用。我們提出數種效能評估方式，並與其他的物件偵測演算法比較（例如: faster R-CNN, YOLO 等）。提出的方法都顯現出優異的效能。我們也希望研究社群，可以本於此研究以及資料集，繼續發覺無人機的優異效能。
我們得感謝聯發科 MediaTek 這幾年來大聯盟計畫上實質的協助跟研究方向的支持；也感謝 NVIDIA在研究經費以及GPU設備上的大力相助。
Demo video:
[1] CARPK dataset: https://lafi.github.io/LPN/
[2] Meng-Ru Hsieh, Yenliang Lin, Winston H. Hsu. Drone-based Object Counting by Spatially Regularized Regional Proposal Networks. ICCV 2017. (PDF)