What is this?¶
Predict depth of transparent object in pixel-wise with Fully Convolutional Networks.
Raw RGB image.
Raw depth image.
Output depth image. The value of each pixel is equal to
~output/depth_pred_rawin the region labeled as transparent, otherwise equal to
~input/depth. The image encoding is 32FC1.
Output label image. Each object is segmented according to param
Probability image of each object predicted according to param
~target_names. If the number of classes including background is XX, then the image encoding is 32FCXX.
Predicted whole depth image. This is used for generating
~output. The image encoding is 32FC1.
Framework for neural networks. Currently, only
-1represents CPU mode.
~target_names(List of String, Required)
Target names for classification.
Saved .npz file for trained model.
Label value for background. This is used with rosparam
Threshold for labeling pixels as uncertain, and the uncertain region will be labeled as background with rosparam
How many messages you allow about the subscriber to keep in the queue. This should be big when there is much difference about delay between two input topics.
Whether to use approximate for input topics.
How many seconds you allow about the difference of timestamp. This is used only when param
roslaunch jsk_perception sample_fcn_depth_prediction.launch