While recent deep learning-based object detection has achieved great success in various fields, it remains challenging to find tiny objects in aerial imagery on-the-fly using mobile devices. Since mobile platforms such as drones operate with limited onboard computing power, handling high-resolution images to find tiny objects with compute-intensive deep learning-based applications often fails to meet their real-time constraints. To mitigate this problem, we propose HashEye, a novel framework that enables fast on-drone tiny object detection by efficiently suppressing spatial redundancy in aerial imagery. HashEye utilizes a lightweight hashing algorithm to rapidly scan image patches; patches exhibiting high hash collision frequencies are identified as background and suppressed. Subsequently, the remaining salient patches are dynamically rearranged into a hardware-friendly dense image for efficient inference. Experimental results on two real-world datasets demonstrate that HashEye achieves up to a 5.25× speedup compared to the baseline, maintaining detection capability.
Hong et al. (Wed,) studied this question.