Falling objects from buildings can cause severe injuries to pedestrians due to the great impact force they exert. Although surveillance cameras have been mounted around some buildings, it is difficult for humans to spot such events by watching surveillance videos due to the small sizes and fast motion of falling objects and complex background. It is very necessary to develop algorithms to automatically detect the falling objects around buildings in surveillance videos. To facilitate the development of falling object detection algorithms, we propose a large diverse video dataset, termed FADE, for FAlling object DEtection around buildings. Our dataset contains 1881 videos taken from 18 scenes with 8 falling object categories, weather conditions and 4 video resolutions. We also propose a new object detection method which effectively leverages motion information for falling object detection around buildings. We extensively evaluate and analyze our method and other 13 previous methods including generic object detection methods, video object detection methods and moving object detections on the proposed FADE dataset. Our method achieves significant improvements over the other methods, providing an effective baseline for future research.
• videos in diverse weather conditions, light conditions, and scenes.
• 8 kinds of objects as follows: clothes, shoes, kitchen waste, books, spitballs, bottles, packaging
bags, and packaging boxes.
• three different camera angles: 30°, 45°, and 60°.
• videos with four resolutions: 1,280 × 720、 1,920 × 1,080、2,560 × 1,440 and 2,592 × 1,520.
• data with diverse backgrounds.
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.