search

makefile / frcnn

Faster R-CNN / R-FCN :bulb: C++ version based on Caffe
Other
184 stars 75 forks source link

How to train on coco? #14

Open lucasjinreal opened 5 years ago

lucasjinreal commented 5 years ago

How to train on coco anyway?

makefile commented 5 years ago

The input data format this repo support is as follows:

# image-id
image-name
number of boxes
label x1 y1 x2 y2 difficulty
...

So if you like to use the input layer in this repo, you need convert coco to this format, or you can custom your input layer.

lucasjinreal commented 5 years ago

Oh, I see....

lucasjinreal commented 5 years ago

@makefile BTW, how to design anchor in terms of coco dataset? For those voc it's like this:

 "feat_stride": "16",
 "anchors": "-84, -40, 99, 55,
             -176,  -88,  191,   103,
             -360,  -184,   375,   100,
             -56,   -56,    71,    72,
             -120,  -120,   135,   135,
             -248,  -248,   263,   263,
             -36,   -80,    51,    95,
             -80,   -168,   95,    183,
             -168,  -344,   183,   359",
 "test_score_thresh": "0.5",
 "n_classes": "21",

How to generats anchors in terms of coco dataset?

makefile commented 5 years ago

Try to use examples/FRCNN/generate_anchors.py to generate anchors info.

lucasjinreal commented 5 years ago

@makefile Should I change anything in terms of different dataset? I can not found any where to set if I am not using voc.

makefile commented 5 years ago

You can adjust the settings in voc cfg file to adapt to different dataset.

lucasjinreal commented 5 years ago

I mean, should those anchors also changed? But I can not see any params to set in generate_anchors.py.

makefile commented 5 years ago

You can set params in function generate_anchors(base_size=16, ratios=[0.5, 1, 2], scales=2**np.arange(3, 6)).