Detectron2

Detectron2 is Facebook AI Research's next generation software system that implements state-of-the-art object detection algorithms. It is a ground-up rewrite of the previous version, Detectron, and it originates from maskrcnn-benchmark. It consists of:

Training recipes for object detection, instance segmentation, panoptic segmentation, semantic segmentation and keypoint detection.
80+ pre-trained models to use for fine-tuning (or training afresh).
Dataset support for popular vision datasets such as COCO, Cityscapes, LVIS, PASCAL VOC, ADE20k.

Tasks

Choose a task to see what models are available:

Object Detection

Semantic Segmentation

Instance Segmentation

Panoptic Segmentation

Models

Viewing Models for Object Detection:

MODEL	TRAINED ON	BOX AP	~FLOPS	YEAR
TensorMask	COCO	41.4	504 Billion	2019
TridentNet	COCO	43.6	888 Billion	2019
RetinaNet	COCO	40.4	273 Billion	2017
Mask R-CNN	COCO	44.3		2017
Faster R-CNN	COCO	43.0	406 Billion	2015
Fast R-CNN	COCO	37.8		2015

Models

Viewing Models for Semantic Segmentation:

MODEL	TRAINED ON	MIOU	PARAMETERS	YEAR
PointRend	Cityscapes	78.9	48 Million	2019
DeepLabV3+	Cityscapes	80.0	60 Million	2018
DeepLabV3	Cityscapes	78.5	58 Million	2017

Models

Viewing Models for Instance Segmentation:

MODEL	TRAINED ON	MASK AP	~FLOPS	YEAR
PointRend	COCO	41.1		2019
TensorMask	COCO	35.8	504 Billion	2019
Mask R-CNN	COCO	39.5		2017

Models

Viewing Models for Panoptic Segmentation:

	MODEL	TRAINED ON	PQ	PARAMETERS	PAPER	YEAR
	Panoptic-DeepLab	COCO	35.5	30 Million		2019
	Panoptic FPN	COCO	43.0	65 Million		2019

facebookresearch / detectron2

FAIR's platform for object detection and semantic segmentation

Detectron2

Tasks

Object Detection

Semantic Segmentation

Instance Segmentation

Panoptic Segmentation

Models

Models

Models

Models