Semi-supervised learning method for object detection in autonomous vehicle and server for performing semi-supervised learning for object detection in autonomous vehicle
A semi-supervised learning method for object detection in an autonomous vehicle and a device for performing semi-supervised learning for object detection in an autonomous vehicle can include receiving, by a server, no-label voxel data from a vehicle, performing, by the server, a data-based update on a server object detection model on the basis of label voxel data and the no-label voxel data, determining, by the server, a loss value on the basis of the label voxel data and the no-label voxel data, and performing, by the server, a loss-based update on the server object detection model using the loss value.
1. A semi-supervised learning method for object detection in an autonomous vehicle, the semi-supervised learning method comprising:
receiving, by a server, no-label voxel data from a vehicle;
performing, by the server, a data-based update on a server object detection model on the basis of label voxel data and the no-label voxel data;
determining, by the server, a loss value on the basis of the label voxel data and the no-label voxel data; and
performing, by the server, a loss-based update on the server object detection model using the loss value,
wherein the loss value includes a supervised loss and a consistency loss,
the supervised loss is determined on the basis of a determination result obtained by inputting object information included in the label voxel data and the label voxel data into the server object detection model,
the consistency loss includes a first consistency loss based on a no-label determination result and a second consistency loss based on a label determination result,
the first consistency loss is determined on the basis of a first sub-no-label determination result obtained by inputting the no-label voxel data into the server object detection model and a second sub-no-label determination result obtained by inputting the image-processed no-label voxel data into the server object detection model, and
the second consistency loss is determined on the basis of a first sub-label determination result obtained by inputting the label voxel data into the server object detection model and a second sub-label determination result obtained by inputting the image-processed label voxel data into the server object detection model.
2. The semi-supervised learning method of claim 1 , wherein the data-based update is performed on the basis of the label voxel data, image-processed label voxel data obtained by performing image processing on the label voxel data, the no-label voxel data, and image processed no-label voxel data obtained by performing image processing on the no-label voxel data, and
the image-processed label voxel data and the image-processed no-label voxel data are generated through image processing based on a travel direction of the vehicle.
3. A server for performing semi-supervised learning for object detection in an autonomous vehicle, the server comprising:
a communicator configured to receive no-label voxel data from a vehicle; and
a processor configured to perform a data-based update on a server object detection model on the basis of label voxel data and the no-label voxel data, determine a loss value on the basis of the label voxel data and the no-label voxel data, and perform a loss-based update on the server object detection model using the loss value,
wherein the loss value includes a supervised loss and a consistency loss,
the supervised loss is determined on the basis of a determination result obtained by inputting object information included in the label voxel data and the label voxel data into the server object detection model,
the consistency loss includes a first consistency loss based on a no-label determination result and a second consistency loss based on a label determination result,
the first consistency loss is determined on the basis of a first sub-no-label determination result obtained by inputting the no-label voxel data into the server object detection model and a second sub-no-label determination result obtained by inputting the image-processed no-label voxel data into the server object detection model, and
the second consistency loss is determined on the basis of a first sub-label determination result obtained by inputting the label voxel data into the server object detection model and a second sub-label determination result obtained by inputting the image-processed label voxel data into the server object detection model.
4. The server of claim 3 , wherein the data-based update is performed on the basis of the label voxel data, image-processed label voxel data obtained by performing image processing on the label voxel data, the no-label voxel data, and image-processed no-label voxel data obtained by performing image-processing on the no-label voxel data, and
the image-processed label voxel data and the image-processed no-label voxel data are generated through image processing based on a travel direction of the vehicle.