kitti-semseg-trainwith annotated data and
kitti-semseg-testwith test images.
kitti-veh-tr. We will merge a few classes (different vehicles) into a single one, drop objects of other classes and do special image augmentations to simulate image scaling we will use during inference.
unet-kitti-vehmodel to test images and analyse the results.
kitti-semseg-trainproject -> Statistics -> Objects area. Supervisely automatically calculates all statistics in real time and it allows us to get valuable insigts from our data. Then we can use this information to create the "right" training set. This dataset contains 34 classes. Let's analyze the classes that have the largest area.
background) we will face the class imbalance problem. To deal with this we will use UNet V2 architecture with additional Dice loss function (it turns on automatically).
kitti-semseg-train(200 images) and create a new one -
kitti-veh-tr(2400 images). A few interesting comments:
"action": "data"). We map classes
trailerto a single class
vehicle. Other classes will be dropped (because of this field :
unet-kitti-veh. The project
kitti-veh-tris used for training.
"val_every": 0.5in the training config). You can find here how to do it. We assigned the
unet-kitti-veh (ckpt 27)name to this restored model.
unet-kitti-veh (ckpt 27)to the project
kitti-semseg-test. The resulting project with neural network predictions will be saved as
"save_classes": [ "vehicle" ]just to automatically drop background object after inference.