[question] any instruction on train multi custom classes?

[eisneim]

Hi, suppose i wanna use kittiBox to detect road-sign, billboard, monitor screen and totally 7 types of object, i seem like i have to modify AnnotationLib.py to make it parse annotation rect with class number.

thanks!

[MarvinTeichmann]

Hi,

yes, you do. In addition you have to modify the loss. Currently it does this:

true_classes = tf.reshape(tf.cast(tf.greater(classes, 0), 'int64'), [outer_ size * hypes['rnn_len']])

This is implementation is actually quite ugly and only works for two classes. Make sure that you generate a vector representing class labels as one-hot vector or as numerical class vector. You can then use this vector in combination with: tf.nn.sparse_softmax_cross_entropy_with_logits or tf.nn.softmax_cross_entropy_with_logits depending on your vector encoding.

Btw. I am also going to try to reduce the annonlib dependency. So you might want to keep an eye on the next one or two commits.

[eisneim]

@MarvinTeichmann great! I'll give it a try

[MarvinTeichmann]

As mentioned in the commit messages, keep a close look at the modifications of fab6bae and bb25fc6. The functions in fab6bae can help you to visualize what is going on during data loading. bb25fc6 removes most of the dependency to AnnoLib. Data loading is now done directly fromt he kitti files. You might want to do somethink along the lines of bb25fc6 for you stuff.

[eisneim]

@MarvinTeichmann wow, thank you so much

[eisneim]

@MarvinTeichmann after spending few hours reading the code, i just got some basic understanding of it, and there is an error initializing tf computation graph:

File "/Users/eisneim/www/deepLearning/KittiBox/hypes/../decoder/fastBox.py", line 382, in loss
    name='reg_loss')
File "/Library/Python/2.7/site-packages/tensorflow/python/ops/math_ops.py", line 1951, in add_n
    raise ValueError("inputs must be a list of at least one Tensor with the "
ValueError: inputs must be a list of at least one Tensor with the same dtype and shape

since YOLO v2 is working OK for my project, so i guess i'll try to fine-tune YOLO model first and keeping an eye on KittiBox, definitely gonna test out KittiBox later
Thanks

[MarvinTeichmann]

Did you run git submodule update after your last pull? The reason for the error is, that you are using an old tensorflow-fcn version together with the new kittibox version. (Or vice versa).

[eisneim]

oops! that's my mistake, issue fixed

[MarvinTeichmann]

@eisneim may I asked which Yolo implementation you are using? I would like to add another baseline to the paper.

[eisneim]

@MarvinTeichmann darknet for training (using YOLOv2 network) use tensorflow implementation darkflow for other stuff, use darknet binary I can get about 30-40fps performance, but accuracy varies across different class due to training datasets(which is hand annotated using my chrome extension)

my ultimate goal is to not only detect the object, but also estimate object's surface homography(4 points) so i can replace and render some other image on top of detected surface

[yangyuekai]

Hi, I want to know that did you change the code to multi class? If you did,can you give me a instruction? I have no idea on the change of code. Thanks a lot if you can answer me. It's better if you can help me.