Research · Jun 2026
Self-Supervised Video Object Segmentation for Autonomous Driving: Cross-Domain Evaluation on KITTI and Waymo
Cross-domain evaluation of self-supervised video object segmentation for autonomous driving.
This paper looks at how well SSL-UVOS, a self-supervised video object segmentation method built on a frozen DINO ViT-S/8 backbone, actually performs when applied to real-world driving data. Self-supervised approaches are appealing because they skip the expensive labeling process that supervised methods depend on, but nobody had really tested how these methods hold up once you move them from clean benchmark datasets into messier, more complex driving scenes.
We ran twelve pipelines that varied the pretraining domain, the evaluation dataset, and the input resolution, testing everything on two major driving benchmarks: KITTI MOTS and the Waymo Open Dataset. For some pipelines we also fine-tuned the model using SimCLR contrastive pretraining on unlabeled driving footage to see if domain-specific adaptation would help. Performance was measured using FG-ARI and J&F, and we paid particular attention to far-zone segmentation, since detecting small, distant objects like pedestrians and cyclists early is one of the most safety-critical parts of the problem.
The results show a clear gap: SSL-UVOS performs noticeably better on KITTI than on Waymo, likely because Waymo scenes are more complex and further removed from the kind of data the underlying DINO backbone was originally trained on. Bumping up the input resolution turned out to be the single most reliable way to improve far-zone segmentation on KITTI. Domain-specific pretraining, on the other hand, barely moved the needle and sometimes made things worse, especially on Waymo. Overall, the work points to resolution, not contrastive pretraining, as the more practical lever for improving detection of small, distant objects in real driving conditions, while also showing just how far self-supervised methods still have to go before they're ready for safety-critical autonomous driving perception.
- Computer Vision
- Deep Learning
- Machine Learning