maybe a bug!!! about face-magnet HOT 4 OPEN

Hi @foralliance
Thanks for pointing these out. It's nothing to be worried about as a major bug. csvdb was something that we developed internally to read the datasets from csv files.
I updated these points, please let us know if you find more.

from face-magnet.

@po0ya HI
Another two questions:

There are three corresponding xxxx_multi layers under the RPN folder. In the prototxt file of the models/wider/VGG16/train folder, you don't use these layers.
By analyzing the code, i think the three layers should correspond to the case of IMS_PER_BATCH >1. right???

Take facemagnet.prototxt for example, I replaced the AnchorTargetLayer, ProposalLayer, ProposalTargetLayer, with the corresponding xxx_multi layer, then modified IMS_PER_BATCH=2.
I found that the code could also be trained properly.
Does this indicate that the code in this article supports IMS_PER_BATCH>1? .In py-faster-rcnn, it seems that only IMS_PER_BATCH=1 is supported.

that's my understanding, I'm not sure it is right.

Can you provide a trained model for debugging test code.

from face-magnet.

from face-magnet.

@po0ya HI
Another three questions:

In skipface.prototxt(train), in RPN NORM section, the Deconvolution layer's output dimension is 1.
Is there any advantage in using settings where the output dimension is 1????

For PASCAL Faces detection, can you provide the relevant pascal_imglist.csv file????

The test results of FDDB data are all rectangular boxes, like:
195 62 128 204 0.966717064381
but it seems the ellipse box(58.984614 39.593000 -1.345725 115.094800 81.162477 1) is required when the official test is provided. how to transfer the rectangle box to ellipse box???

Another two little problems:

In readme, in the Single image detection section, the example you gave is

. / experiments/scripts/test_single.sh facemagnet data/demo/demo0.jpg --pyramid

I think we should add a parameter max_size, that is

. / experiments/scripts/test_single.sh facemagnet data/demo/demo0.jpg --max_size 3000 --pyramid

Because the original cfg.TEST.MAX_SIZE = 1500 was too small to to reflect the pyramid advantage, especially when scale is equal to 2

When testing the FDDB dataset, according parameters: args.single=False, imdb='fddb', an error is reported when output_dir is obtained．

Traceback (most recent call last):
  File "./tools/test_net.py", line 171, in <module>
    output_dir = get_output_dir(imdb, postfix=postfix)
  File "/home/jmx/face-magnet/tools/../lib/fast_rcnn/config.py", line 236, in get_output_dir
    outdir = osp.abspath(osp.join(__C.ROOT_DIR, 'output', __C.EXP_DIR, imdb.name))
AttributeError: 'str' object has no attribute 'name'

However, this error does not affect the final test

from face-magnet.