search

matejgrcic / DenseHybrid

Official implementation of paper "DenseHybrid: Hybrid Anomaly Detection for Dense Open-set Recognition"
GNU General Public License v2.0
36 stars 2 forks source link

Question: What class are the pasted anomalies considered as? #4

Closed mbisan closed 1 year ago

mbisan commented 1 year ago

In line 115 of the train function the labels with anomalies marked as self.args.num_classes + 1 is marked on those positions as self.args.num_classes. For the cityscapes dataset this would mean setting the class at the pasted positions to 19, which is class unlabeled, or rather the class ignored when computing the cost.

Is this right?

matejgrcic commented 1 year ago

The classification loss is not computed in the pasted negative object. We implement this by assigning ignore label to the corresponding pixel.

mbisan commented 1 year ago

Ok, thank you for confirming this.