dataset.md

Datasets & Benchmarks

Detection

The standard OV-Det use OV-COCO and OV-LVIS datasets for training and evaluation. The detectors are trained on base classes and test both base and novel classes for evaluation.

Datasets	base	novel	Tasks
COCO OVD	48	17	det.
LVIS OVD	866	337	det.

Segmentation

The OVSS task can be evaluated with different protocols (self-evaluation, cross-evaluation). In general, there is not a consensus on evaluation protocol like OVOD. Please see Sec. 4.2 for details.

Open Vocabulary Semantic Segmentation on self-evaluation setting

Datasets	base	novel
COCO-Stuff	156	15
PASCAL VOC	15	5
PASCAL Context	49	10

Open Vocabulary Semantic Segmentation on cross-evaluation setting

Datasets	base	novel
A-847 (ADE20K)	-	847
PC-459 (PASCAL Context)	-	459
A-150 (ADE20K)	-	150
PC-59 (PASCAL Context)	-	59
PS-20 (PASCAL VOC)	-	20

Open Vocabulary Instance Segmentation

Datasets	base	novel
COCO-Instances	48	17

Open Vocabulary Instance Segmentation

Datasets	base	novel
COCO-Instances	48	17

Open Vocabulary Panoptic Segmentation

Datasets	base	novel
ADE20K	135	15
COCO-Panoptic	156	15