tobiasfshr / motsfusion Goto Github PK

This is a very interesting work and I tried to run this, but some constructions are deprecated and don`t work now.
At first, for user needs to setup his environment:
I used anaconda, in this case, i can offer this way of setup(this instruction will safe a lot of time for other users, moreover it will be helpfull make this in sh script)

conda create --name=labelme python=3.6
source activate labelme

or deprecated source /opt/anaconda/bin/activate labelme

after this, it needs to add channels

conda config --add channels conda-forge hcc omnia menpo

then, install libraries. Attention, this work was written by tensorflow 1 and it needs to disable TFv2 behavior or install TFv1 in your environment

conda install numpy, opencv, scipy, pycocotools, tensorflow=1.15, munkres, imbalanced-learn, vtk, pandas, scikit-image

First task in the setup environment is over at this moment. Next, it needs to create structure of folders with images, detectors, models, etc. In readmy this part was not written obviously.

At first, from data_tracking_calib dataset files needs to put in ./data/KITTI_tracking/train/images/*.txt
At second, put BBSegNet files in .external/BBSegNet.
And put detections in ./ of project(actually)

I started precompute.py --config=/configs/config_default and python read detectorts,

Could you tell me more about structure of folders? It seems python couldn`t find some images(or instances in imgs).

P.S.:If you interesting in, i could get more info from clean instalation

P.P.S.: and I needed to change matplotlib.misc to scipy.misk because it was replaced too

Hi,
could you publish your validation 3D MOT data (i.e. with 3D OOBBs)?

Otherwise, could you give step-by-step instructions on running your model so that I can create the results myself (e.g. how to compute optical flow etc)?

Thanks!

Hi,

Thanks for your wonderful work. Could you please release the 3D tracking and reconstruction results produced by your model?

This snippet is taken from the BB2Net xception.py module:

    def _separable_conv(features, depth, kernel_size, depth_multiplier,
                        regularize_depthwise, rate, stride, scope):
      if activation_fn_in_separable_conv:
        activation_fn = tf.nn.relu
      else:
        activation_fn = None
        features = tf.nn.relu(features)
      return separable_conv2d_same(features,
                                   depth,
                                   kernel_size,
                                   depth_multiplier=depth_multiplier,
                                   stride=stride,
                                   rate=rate,
                                   activation_fn=activation_fn,
                                   regularize_depthwise=regularize_depthwise,
                                   scope=scope)
    for i in range(3):
      residual = _separable_conv(residual,
                                 depth_list[i],
                                 kernel_size=3,
                                 depth_multiplier=1,
                                 regularize_depthwise=regularize_depthwise,
                                 rate=rate*unit_rate_list[i],
                                 stride=stride if i == 2 else 1,
                                 scope='separable_conv' + str(i+1))
    if skip_connection_type == 'conv':
      shortcut = tf.Conv2D(inputs,
                             depth_list[-1],
                             [1, 1],
                             stride=stride,
                             activation_fn=None,
                             scope='shortcut')
      outputs = residual + shortcut
    elif skip_connection_type == 'sum':
      outputs = residual + inputs
    elif skip_connection_type == 'none':
      outputs = residual
    else:
      raise ValueError('Unsupported skip connection type.')

    return slim.utils.collect_named_outputs(outputs_collections,
                                            sc.name,
                                            outputs)

In the last line, we're using the slim module from tf.contrib, which is deprecated in tensorflow 2. What functions exist in tensorflow 2 or otherwise to do the same function as the slim.utils.collect_named_outputs line?