Due to the increase in computational resources and accessibility of data, an increase in large, deep learning models trained on copious amounts of data using self-supervised or semi-supervised learning have emerged. These foundation models are often adapted to a variety of downstream tasks like classification, object detection, and segmentation with little-to-no training on the target dataset. In this work, we perform a robustness analysis of Visual Foundation Models (VFMs) for segmentation tasks and compare them to supervised models of smaller scale. We focus on robustness against real-world distribution shift perturbations. We benchmark four state-of-the-art segmentation architectures using 2 different datasets, COCO and ADE20K, with 17 different perturbations with 5 severity levels each. We find interesting insights that include (1) VFMs are not robust to compression-based corruptions, (2) while the selected VFMs do not significantly outperform or exhibit more robustness compared to non-VFM models, they remain competitively robust in zero-shot evaluations, particularly when non-VFM are under supervision and (3) selected VFMs demonstrate greater resilience to specific categories of objects, likely due to their open-vocabulary training paradigm, a feature that non-VFM models typically lack. We posit that the suggested robustness evaluation introduces new requirements for foundational models, thus sparking further research to enhance their performance.
Visual example from the COCO-P dataset for JPEG compression under varying levels of severity. The left shows results for ODISE, middle shows the original images, and the right shows GroundedSAM. We see as severity increases, both models mask quality decreases but ODISE additionally incorrectly classifies objects, such as “person” to “oven”
The data for COCO-P and ADE20K-P are available for download here.
Directory Structure:
|-- <corruption>
| |-- <severity>
| | `-- coco
| | |-- annotations
| | | |-- captions_val2017.json
| | | |-- instances_val2017.json
| | | |-- panoptic_caption_val2017.json
| | | |-- panoptic_val2017.json
| | | `-- person_keypoints_val2017.json
| | |-- panoptic_semseg_val2017
| | |-- panoptic_val2017
| | `-- val2017Directory Structure:
|-- <corruption>
| `-- <severity>
| `-- ade
| `-- ADEChallengeData2016
| |-- ade20k_instance_val.json
| |-- ade20k_panoptic_val
| |-- ade20k_panoptic_val.json
| |-- annotations
| | `-- validation
| |-- annotations_detectron2
| | `-- validation
| |-- annotations_instance
| | `-- validation
| |-- images
| | `-- validation
| |-- objectInfo150.txt
| `-- sceneCategories.txt
COCO is under Creative Commons Attribution 4.0 and ADE20K is under BSD 3-Clause License. Please see COCO and ADE20K for more details. Licensing information for each model are located in models/.
See README.md for installation instructions and model details.
Our work is relies on the open-source models built and inspired by several assets. We gratefully thank the authors for their open-source projects that allowed this benchmark to be possible:
- ODISE: https://github.com/NVlabs/ODISE/
- Mask2Former: https://github.com/facebookresearch/Mask2Former/
- MaskDINO: https://github.com/IDEA-Research/MaskDINO
- Detectron2: https://github.com/facebookresearch/detectron2/
- mmsegmentation: https://github.com/open-mmlab/mmsegmentation
- COCO panoptic: https://github.com/cocodataset/panopticapi