we construct a benchmark video dataset termed as FADE, the first one for FAlling object DEtection around buildings. Our dataset is diverse, which contains nine object categories from sixteen scenes with three weather conditions, and four video resolutions. Since falling object detection 13 around buildings (FODB) can be considered as a special moving object detection 14 (MOD) task, in addition to using the evaluation metrics of MOD, we propose a new 15 evaluation metric named TRO to measure the performance of the methods to locate 16 the beginning and ending time of the falling incident

Falling objects (even small ones) from the high-rise buildings, which have the characteristics of the fast moving speed and short collision time, would cause severe harm to the people who pass by, as the falling objects can produce tremendous impact force. However, due to the reasons that the background is complicated and changeable (e.g. illumination and weather variation), falling objects in the monitoring video are small and sometimes blurred due to fast motion, and there is no available benchmark dataset specialized for falling object detection (the most crucial factor), the accuracy of the existing falling object detection methods is low.

we propose a new dataset termed as FADE, which contains 1,189 videos, i.e., 298,810 video frames, of which 50,472 frames are finely annotated covering various scenes, weather conditions, light conditions, and video resolutions. Different from the existing MOD datasets, FADE is the first dataset specialized for falling object detection around the building. To evaluate the FODB algorithms more effectively, we design a new evaluation metric named TRO, which aims at measuring the ability of the approaches in locating the beginning and ending time of the falling incident. Besides, we provide a benchmark for the existing popular MOD methods. The results show that FODB is a challenging task under the interference of complex backgrounds and motion blur, and our FADE is a heuristic dataset which supplies realistic data to promote the progress of FODB.



• We construct a new video dataset named FADE, the first dataset used for detecting falling objects around the building. It is diversity, which covers various scenes and complex conditions, e.g. changeable weather, alterative light, different camera angles, etc.
• We evaluate 10 MOD methods comprehensively, which can be used as a benchmark for future research on FODB.
• We propose a new evaluation metric TRO, which assesses the temporal localization capacity of the FODB method.


      License: Our FADE dataset is published under the CC BY-NC-SA 4.0 license. Our code is released under the Apache 2.0 license.