From this line, I assume you use RGB: img = img - np.array([104,117,123])

However opencv uses BGR, so this line returns BGR: if args.path=='CAMERA': ret, img = cap.read()

Then BGR is fed to the network bboxlist = detect(net,img)

I fed RGB to the network and got worse results. Is it possible that you meant RGB in all places but the network is actually trained for BGR? (If then it should be img = img - np.array([123,117,104]))

Dear @clcarwin, Thank you for your nice work. Would you please tell me how you can convert Caffe weights and model of S3FD into PyTorch? Can you convert the model & pre-trained weights of RefineDet into PyTorch?

hi @clcarwin,

I test you evaluation results on wider face as (easy 92.8, medium 91.5, hard 84.2). But with the original model provided by sfzhang15/SFD, I can get (easy 93.8, medium 92.4, hard 85.1).

Did I test correctly? If so, why there is accuracy loss?

Great work! Best,

Sir,I'm sorry to disturb you about this object. I run this object on windows 10,python 3.5.2 ,pytorch 0.3. After : python test.py --model data/s3fd_convert.pth --path data/test01.jpg, the screen display: D:\Python\Pytorch_cw_sfd\SFD_pytorch>python test.py --model data/s3fd_convert.pth --path data/test01.jpg Traceback (most recent call last): File "test.py", line 71, in bboxlist = detect(net,img) File "test.py", line 27, in detect for i in range(len(olist)/2): olist[i2] = F.softmax(olist[i2]) TypeError: 'float' object cannot be interpreted as an integer

Why ???

Hi @clcarwin,

Why do you set the padding size of fc6 to 3? This is inconsistent with the original paper. See https://github.com/clcarwin/SFD_pytorch/blob/master/net_s3fd.py#L42

Best,

Good: It is accurate.

Bad: The inference time is more than 80 ms for realtime usage. To make it work for realtime image has to be resized to less than 200x200 which reduces accuracy.

So in order to make it usable the only way is to make it faster. Have you tried using TensorRT or TVM or Pytorch serving in C++ ?

Seems nobody's looking at PR's here, but letting others know I've made a number of improvements.

It runs smoothly on modern pytorch (1.3) and refactored the code to eliminate redundant code. I also added some convenient methods that make it easier to do common things, like detect_faces. Also, added integration tests.

I independently found the same speed-up as @kir-dan in https://github.com/clcarwin/SFD_pytorch/pull/4 and moved all that code into pytorch instead of numpy, so it can be fully run on GPU.

Hi, I have been running the model using test.py and modified it run multiple files. The GPU memory keeps on increasing,from 3gigs to 9 gigs. Is this due to poor garbage collection?

Dear @clcarwin, If one wants to change the aspect ratio of anchor boxes, must just changed the detect method in test.py? For example, line https://github.com/clcarwin/SFD_pytorch/blob/96fdfbe22eef176a04802d915834b82a131a854d/test.py#L39 or other methods moreover must changed?

When I use the Tensorflow to build the project, I have some trouble in data augmentation which describe in the paper. Can you tell the details of the data augmentation or show your data augmentation code to me. Thank you