TaskCo Unified AutonomousDriving

Disjoint-balance setting.Wefurther consider amore difficult setting with the lowest quantityof annotations that corresponds to the scenarios of scarce annotations. Classification-oriented methodsCurrently,thestate-of-the-art classification-oriented pre-training methods are mainly based on contrastive learning and online clustering. Detection-oriented methodsDetCo [19] is specially designed for object detection and enforce contrastive learning between the global image and local patches with multi-level supervision. We select DenseCL [18] for comparison. Some of its findings back up our experiment results in Section 3. In [7], this pre-training paradigm is tailored for downstreamsingle-task learningbut not multi-task learning.