RFCN vs CSD-RFCN on VOC07
Does CSD regularization help RCNN-based models and how much?
Created on May 4|Last edited on May 18
Comment
Below you can turn on and off multiple comparisons via "Run sets". Each run set comprises a baseline trained only on the labeled data and a CSD version with additional regularization and unlabeled data included.
For more details you should check the runs' detail pages by clicking on their names (note: for nicer summaries inside detail pages append user-vlfom to the end of the URL).
Note: EMA smoothing is applied to the plots, real values can be found in "()" brackets by highlighting the point in the plot.
2
2
2
20% voc07 (labeled) + 80% voc07 (unlabeled)
2
Some visualizations from the VOC07 + VOC12 CSD run. Try varying the prob's threshold or checking how the training progresses by changing the step (best to do it after entering the full screen mode, otherwise Wandb may become buggy).
Run set
1
Add a comment