Xudong Wang, Shufan Li, Konstantinos Kallidromitis, Yusuke Kato, Kazuki Kozuka, Trevor Darrell
Open-vocabulary image segmentation aims to partition an image into semantic regions according to arbitrary text descriptions. However, complex visual scenes can be naturally decomposed into simpler parts and abstracted at multiple levels of granularity, introducing inherent segmentation ambiguity. Unlike existing methods that typically sidestep this ambiguity and treat it as an external factor, our approach actively incorporates a hierarchical representation encompassing different semantic-levels into the learning process. We propose a decoupled text-image fusion mechanism and representation learning modules for both "things" and "stuff". Additionally, we systematically examine the differences that exist in the textual and visual features between these types of categories. Our resulting model, named HIPIE, tackles HIerarchical, oPen-vocabulary, and unIvErsal segmentation tasks within a unified framework. Benchmarked on over 40 datasets, e.g., ADE20K, COCO, Pascal-VOC Part, RefCOCO/RefCOCOg, ODinW and SeginW, HIPIE achieves the state-of-the-art results at various levels of image comprehension, including semantic-level (e.g., semantic segmentation), instance-level (e.g., panoptic/referring segmentation and object detection), as well as part-level (e.g., part/subpart segmentation) tasks. Our code is released at https://github.com/berkeley-hipie/HIPIE.
| Task | Dataset | Metric | Value | Model |
|---|---|---|---|---|
| Semantic Segmentation | COCO minival | PQ | 58.1 | HIPIE (ViT-H, single-scale) |
| Semantic Segmentation | COCO minival | mIoU | 66.8 | HIPIE (ViT-H, single-scale) |
| Instance Segmentation | RefCoCo val | Overall IoU | 82.8 | HIPIE |
| Instance Segmentation | RefCOCO+ val | Overall IoU | 73.9 | HIPIE |
| Zero Shot Segmentation | Segmentation in the Wild | Mean AP | 41.6 | HIPIE |
| Referring Expression Segmentation | RefCoCo val | Overall IoU | 82.8 | HIPIE |
| Referring Expression Segmentation | RefCOCO+ val | Overall IoU | 73.9 | HIPIE |
| 2D Semantic Segmentation | Pascal Panoptic Parts | mIoUPartS | 63.8 | HIPIE (ViT-H) |
| 2D Semantic Segmentation | Pascal Panoptic Parts | mIoUPartS | 57.2 | HIPIE (ResNet-50) |
| 10-shot image generation | COCO minival | PQ | 58.1 | HIPIE (ViT-H, single-scale) |
| 10-shot image generation | COCO minival | mIoU | 66.8 | HIPIE (ViT-H, single-scale) |
| Panoptic Segmentation | COCO minival | PQ | 58.1 | HIPIE (ViT-H, single-scale) |
| Panoptic Segmentation | COCO minival | mIoU | 66.8 | HIPIE (ViT-H, single-scale) |
| Image Segmentation | Pascal Panoptic Parts | mIoUPartS | 63.8 | HIPIE (ViT-H) |
| Image Segmentation | Pascal Panoptic Parts | mIoUPartS | 57.2 | HIPIE (ResNet-50) |