YOLOv7: YOLO with Transformers and Instance Segmentation, with TensorRT acceleration!


YOLOv7: YOLO with Transformers and Instance Segmentation, with TensorRT acceleration!

GitHub – jinfagang/yolov7: ūüĒ•ūüĒ•ūüĒ•ūüĒ• YOLO with Transformers and Instance Segmentation, with TensorRT acceleration! ūüĒ•ūüĒ•ūüĒ•

“In short: YOLOv7 added instance segmentation to YOLO arch. Also many transformer backbones, archs included. If you look carefully, you’ll find our ultimate vision is to¬†make YOLO great again¬†by the power of¬†transformers, as well as¬†multi-tasks training. YOLOv7 achieves mAP 43, AP-s exceed MaskRCNN by 10 with a convnext-tiny backbone while simillar speed with YOLOX-s, more models listed below, it’s more accurate and even more lighter!

Just another yolo variant implemented based on detectron2. But note that¬†YOLOv7 doesn’t meant to be a successor of yolo family, 7 is just a magic and lucky number. Instead, YOLOv7 extend yolo into many other vision tasks, such as instance segmentation, one-stage keypoints detection etc..

The supported matrix in YOLOv7 are:

  • ¬†YOLOv4 contained with CSP-Darknet53;
  • ¬†YOLOv7 arch with resnets backbone;
  • ¬†YOLOv7 arch with resnet-vd backbone (likely as PP-YOLO), deformable conv, Mish etc;
  • ¬†GridMask augmentation from PP-YOLO included;
  • ¬†Mosiac transform supported with a custom datasetmapper;
  • ¬†YOLOv7 arch Swin-Transformer support (higher accuracy but lower speed);
  • ¬†YOLOv7 arch Efficientnet + BiFPN;
  • ¬†YOLOv5 style positive samples selection, new coordinates coding style;
  • ¬†RandomColorDistortion, RandomExpand, RandomCrop, RandomFlip;
  • ¬†CIoU loss (DIoU, GIoU) and label smoothing (from YOLOv5 & YOLOv4);
  • ¬†YOLOF also included;
  • ¬†YOLOv7 Res2net + FPN supported;
  • ¬†Pyramid Vision Transformer v2 (PVTv2) supported;
  • ¬†WBF (Weighted Box Fusion), this works better than NMS,¬†link;
  • ¬†YOLOX like head design and anchor design, also training support;
  • ¬†YOLOX s,m,l backbone and PAFPN added, we have a new combination of YOLOX backbone and pafpn;
  • ¬†YOLOv7 with Res2Net-v1d backbone, we¬†found res2net-v1d¬†have a better accuracy then darknet53;
  • ¬†Added PPYOLOv2 PAN neck with SPP and dropblock;
  • ¬†YOLOX arch added, now you can train YOLOX model (anchor free yolo) as well;
  • ¬†DETR: transformer based detection model and¬†onnx export supported, as well as TensorRT acceleration;
  • ¬†AnchorDETR: Faster converge version of detr, now supported!
  • ¬†Almost all models can export to onnx;
  • ¬†Supports TensorRT deployment for DETR and other transformer models;
  • ¬†It will integrate with¬†wanwu, a torch-free deploy framework run fastest on your target platform…”

Source: github.com/jinfagang/yolov7

June 29, 2022
Notify of
Inline Feedbacks
View all comments

Subscribe to our Digest