Our task is to do a special object detection: In the traditional case, the neural network will output some rectangle bounding boxes. But in our case, the network should output many nearly-vertical lines. The lines will cut the image into many pieces, each piece will be one object. (Our object distribution is very special, and such lines perfectly exist.)
EDIT - My naive idea: Let the input image be 512x512x3. The output will be 1x64x3. The 3 channels are (a) the y coordinate of the line when x=256 (b) the angle of line (c) the confidence that the line exists. Note that the 64x1 means this output is very fat. Actually, it is just like YOLO/SSD, but in our case each row predicts one line, instead of each cell predicting one box. However, I really have no idea how to implement the inner network structure... I would appreciate for any ideas!