Closed anonym24 closed 6 years ago
The input size is 416x416 because it gets scaled down 5 times to be 13x13. If you set the input to 256x256 and scale it down 5 times the final grid becomes 8x8, and so you'll have to adjust the code that computes the bounding boxes for that too. (But I'm not sure how well this will actually work...)
of course accuracy will become much worse but it should process faster, thanks for the answer I'll try to change grid values
could you also please tell me on which devices did you tested this project and how fast it works, iPhone 7, 8 or mb even X?
I had time to test only on old iPad mini 2 (2 core, 1.3 GHz) :)
also would be great to implement Tensorflow object model (ssd_mobilenet_v1_coco) for CoreMl
https://github.com/tensorflow/models/tree/master/research/object_detection https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md
There is already a project to run Tensorflow object detection model on iOS, but it works with Tensorflow (and it's Objective-C) https://github.com/JieHe96/iOS_Tensorflow_ObjectDetection_Example
I think this gets about 10 FPS on the iPhone 7. I haven't tried on 8 or X. However, I did implement MobileNet+SSD (for a client, so I cannot share the code) and it gets 30 FPS on iPhone 7. So it's a lot faster than Tiny YOLO, and I think it's similar in accuracy.
wow results sounds good:) yeah I understand you can't share it
thanks for answers
I tried to set smaller values
but the rects aren't drawn correctly