Data collection

We use a wide dynamic range camera, equipped with a 1/3 progressive scanning CMOS sensor and dot matrix LED infrared lamp. The highest output video quality of the camera is 2,592 × 1,520 @ 30 FPS. We throw objects on the high-rise buildings of the school, and then we use the above-described camera to record the whole process of these events. All of our experiments have been approved by the corresponding school.

Data Format

The annotation file keeps consistent with Pascal VOC and contains two keys: "JPEGImages" and "Annotations".

The data organization is listed as follows:

dataset: {

		| - - - Video 1

			| - - - JPEGImages
            			| - - - frame1.jpg
                    
                    		| - - - frame2.jpg
                    
                    		. . .
                    
                    		| - - - frameN.jpg
                
                	| - - - Annotations

				| - - - frame1.xml 

                                            | - - - image size              
                            
                                            | - - - bndbox  
                                                        | - - - xmin 
                        
                                                        | - - - ymin
                        
                                                        | - - - xmax
                        
                                                        | - - - ymax
                                                        
                                 . . .
                                 
                                | - - - frameN.xml 

		| - - - Video 2

		. . .

		| - - - Video 1,881

}

Data Annotation

Different from the pixel-by-pixel annotation method of the MOD datasets , we adopt the annotation manner used in object detection task to generate the bounding box around each falling object from height.

Comparison of the area of the moving object in our dataset and the MOD dataset. (a) The moving object and the GT in the MOD dataset, taking the dataset CDnet 2014 as an example. (b) The moving object in our FADE dataset.

Maintenance

Our data will be maintained for a long time, and we will check the accessibility of our dataset regularly. The first author will provide support.