Closed julingers closed 2 years ago
@joshi-bharat I have solved the problem of bbox_mask. I just used numpy to implement the bboxmask filt in sess.run().
By the way, we can't use if-else structure to define compute graph, but we can use it after sess.run().
@joshi-bharat I just found that it takes more inference time if use numpy. It's more slower than before. So sad.
If the inference is done in the CPU, it should take the same amount of time. I will check this sometime.
I found another problem that the code can't get right bboxmasks to filt x, y, conf. You used predicted box to get bbox_masks in the computer graph. Fist 'if bbox is None' can't be used, for the graph is statics. Second, input value bbox is tensor type,can't use 'for box in bbox' or ' // 'or 'int( )' and other gramma. And How to solve this problem? We need to use Tensor type to get bbox_masks from bbox in the computer graph.