YOLOv7 v1.0 Release!

New features:

• A powerful convnext-tiny driven YOLOv7 model, mAP 43 without any augmentation training;
• Convnext-tiny YOLOv7 (AKA yoloformer) get mAP-s 26 which 10 AP higher than MaskRCNN!;
• YOLO2go proposed, will make a small weight size YOLO model (in mobilenetv3 or efficicientnetv2 backbone), will release in next version.

Beta Release: SparsInst ONNX, SOLOv2, YOLOX-Pose etc.

Since YOLOv7 not just yolo, it aims at a mature modeling tool based on detectron2, so I will continue add more models into it. It now also received a lot of PR from community. Let's build YOLOv7 togather and make a lot of experiment on it.

Current is a Beta release, it becoming close to a Stable release version, we are keep on doing on that.

• YOLOX-Pose model added, now you should able to training YOLOx with Pose head;
• SOLOv2 training support, now you can using YOLOv7 training SOLOv2;
• Sparseinst training support, now you can train sparseinst, we also support ONNX exportation;
• Some bugs fixes in sparseinst;

Feature you will see in the next release:

• Int8 quantization on Sparseinst, CPU Realtime instance segmentation, just from yolov7!
• TensorRT on SOLOv2 and Sparseinst;
• Pretrain model of YOLOX-Pose.

Just start and fork it! Join the community!

Package support, a decent instance segmentation model supported

updates on:

• remove dependencies of mobilecv, which is now optional;
• remove nbnb deps now is optional;
• add setup.py for install, usage can be outscope anywhere and any other project.