fcn_depth_prediction.py¶

What is this?¶

Predict depth of transparent object in pixel-wise with Fully Convolutional Networks.

~output (sensor_msgs/Image)

Output depth image. The value of each pixel is equal to ~output/depth_pred_raw in the region labeled as transparent, otherwise equal to ~input/depth. The image encoding is 32FC1.
~output/label (sensor_msgs/Image)

Output label image. Each object is segmented according to param ~target_names.
~output/proba_image (sensor_msgs/Image)

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.
~output/depth_pred_raw (sensor_msgs/Image)

Predicted whole depth image. This is used for generating ~output. The image encoding is 32FC1.

~backend (String, Default: chainer)

Framework for neural networks. Currently, only chainer is supported.
~gpu (Int, Default: -1)

GPU id. -1 represents CPU mode.
~target_names (List of String, Required)

Target names for classification.
~model_name (String, Required)

Currently, fcn8s_depth_prediction and fcn8s_depth_prediction_concat_first are supported.
~model_file (String, Required)

Saved .npz file for trained model.
~bg_label (Int, default: 0)

Label value for background. This is used with rosparam ~proba_threshold
~proba_threshold (Float, default: 0.0)

Threshold for labeling pixels as uncertain, and the uncertain region will be labeled as background with rosparam ~bg_label.
~queue_size (Int, default: 10)

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.
~approximate_sync (Bool, default: False)

Whether to use approximate for input topics.
~slop (Float, default: 0.1)

How many seconds you allow about the difference of timestamp. This is used only when param ~approximate_sync is true.

roslaunch jsk_perception sample_fcn_depth_prediction.launch