Caffe: implmentation of googlenet SSD

[hengck23]

I wonder if anyone try to use googlenet as feature extraction? I think this is very fast!

This link suggest that googlenet can be used for object detection too.
https://github.com/sanghoon/pva-faster-rcnn

I tried to use googlenet-bn from https://github.com/lim0606/caffe-googlenet-bn for SSD.
I can get speed of at about 22 ms for batch=1 at run time, however the mAP is not high because:

• An object is detected multiple times (gives rise to false positive) because object parts are always detected. For example, an human detection can have 2 detection: the whole body and the head.
  (I note that the ground truth in VOC do consists of different parts, e.g. head half body, full body, legs)
• all trees becomes potted plants!

Some examples of the images are shown below.
How can I resolved such issues? Any help is much appreciated.

[weiliu89]

I tried Inception-v3 a little bit. I could get about 69ish mAP. I converted the TF Inception-v3 model to caffe format, which might not be the right model (e.g. the top1 accuracy on ILSVRC val dataset is lower than the number it should have). I haven't investigated in more details.

The reason that it detects head as human might caused by the data augmentation during training as well.

[shawn-tian]

@hengck23 I'm not sure which layers to use for the mbox prediction. Could you able to share the prototxt of ssd on googlenet? thank you very much

[ngaloppo]

@weiliu89 Would you mind sharing your inception-v3 implementation? Also, have you considered inception-v4? You can generate caffe proto's with the code in this repo: https://github.com/soeaver/caffe-model

[weiliu89]

@ngaloppo I could share a model that I trained long time ago later, however I haven't really optimize for it. Besides, I converted a v3 TF model to caffe format, but I couldn't achieve the same accuracy it should have on CLS-LOC val dataset. Are there converted caffe models for inception v3 and v4?

[ngaloppo]

@weiliu89 Yes, the aforementioned https://github.com/soeaver/caffe-model includes models for inception v3 and v4.

[weiliu89]

@ngaloppo Are there any pretrained models? I only see the scripts defining the networks.

[CHUNYUWANG]

@hengck23 Can you share the prototxt file you used for training googlenet-ssd?

[soeaver]

@weiliu89 I convert the inception-v3 and other networks to caffe, it may be helpful for you.
https://github.com/soeaver/caffe-model/tree/master/cls

[libra7]

@hengck23 Can you share the prototxt file you used for training googlenet-ssd? it may be helpful for us.

[liuguiyangnwpu]

@weiliu89 @ngaloppo can you show me the pycaffe code to create the model about the googlenet or the inception model ? I want to modified those model to detect the small object.
thx !

[liuguiyangnwpu]

I want to know how long have you train the inceptionv3 + SSD model.