kevinlin311tw/cdcl-human-part-segmentation

Repository for Paper: Cross-Domain Complementary Learning Using Pose for Multi-Person Part Segmentation (TCSVT20)

cdcl-human-part-segmentation's Issues

Needed to set "allow_growth"

Just documenting this in case anyone else encounters this issue.
Built the provided dockerfile and ran python3 inference_15parts.py --scale=1 --scale=0.5 --scale=0.75

Saw E tensorflow/stream_executor/cuda/cuda_dnn.cc:373] Could not create cudnn handle: CUDNN_STATUS_INTERNAL_ERROR (among other errors) when calling model.predict.

Resolved the issue by adding the following to inference_15parts.py

import tensorflow as tf
from keras.backend.tensorflow_backend import set_session

config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
set_session(sess)

About the mIOU for CDCL+voc?

I test your provided weight-model_simulated_RGB_mgpu_scaling_append.0024.h5, but i just test the MIOU is 69.53, i am very confused with this, and can you provide your test codes?

training code

Great work！When do you realease training code？

Weights Link Down

Great work!! The model weights link is down... Could you please update it. Thank you.

LICENSE

Hi,

thank you for sharing your work.
Could you add a LICENSE file to the repo?

thanks,

Jeremy

The generation of synthetic pose labels and synthetic part labels

Hi, could you talk about how to get the synthetic pose labels and synthetic part labels of synthetic data, and what software are you used to generate synthetic data? I would like to learn more about the generation of synthetic data.
Thanks a lot.

Full part model

Amazing work. When do you think the pre-trained models with all parts will be available?

inference error

when i run "python3 inference_15parts.py --scale=1" , error happens like:

Traceback (most recent call last):
File "inference_15parts.py", line 200, in
model = get_testing_model_resnet101()
File "D:\body\body\model_simulated_RGB101.py", line 456, in get_testing_model_resnet101
C1, C2, C3, C4, C5 = ResNet101_graph(img_input, None)
File "D:\body\body\model_simulated_RGB101.py", line 250, in ResNet101_graph
x = conv_block(x, 3, [64, 64, 256], stage=2, block='a', strides=(1, 1))
File "D:\body\body\model_simulated_RGB101.py", line 208, in conv_block
x = BatchNormalization(epsilon=eps, axis=bn_axis,name=bn_name_base + '2a')(x)
File "D:\Python3.6\lib\site-packages\keras\engine\base_layer.py", line 463, in call
self.build(unpack_singleton(input_shapes))
File "D:\Python3.6\lib\site-packages\keras\layers\normalization.py", line 99, in build
str(input_shape) + '.')
ValueError: Axis 1 of input tensor should have a defined dimension but the layer received an input with shape (None, None, None, 64).

Has anyone been in the same situation like this?

Inference time

Hi, What is the inference time of the model, On what resolution and on which device has it been tested on? Could you please share these detailes?
Thanks

Can i save one by one?

Like the image

If there are multiple people or overlapping, I want to save each person separately. Is that possible?

Please give me some advice

Any update for training code and synthesis data

I am very interested of this method.
I know it has been a long time that the training code and data were not public because the company did not allow it. So is it approval for now?

CDCL+COCO Model Weights

Hi!

Do you have a release for CDCL+COCO? I would love to use and cite your research but without access to the synthetic training data or model weights I am a bit stuck.

Thank you!

PASCAL-Person-Part-dataset

hi
Thanks for your sharing
Where can I get the PASCAL-Person-Part dataset please?

Synthetic data

Hi,
thanks for sharing the great insight. Taking advantage of synthetic data to promote performance on real data is realy thoughtful!
Many readers are waiting for the scripts to generate the sysnthetic data and also the training code. Please kindly share it so that the community can benefit from your cutting-edge works. Thank you.