YOLOv4-tiny released: 40.2% AP50, 371 FPS (GTX 1080 Ti)

[AlexeyAB]

Discussion: https://www.reddit.com/r/MachineLearning/comments/hu7lyt/p_yolov4tiny_speed_1770_fps_tensorrtbatch4/

YOLOv4-tiny released: 40.2% AP50, 371 FPS (GTX 1080 Ti) / 330 FPS (RTX 2070): https://github.com/AlexeyAB/darknet/issues/6067

• Paper (CVPR 2021): https://openaccess.thecvf.com/content/CVPR2021/html/Wang_Scaled-YOLOv4_Scaling_Cross_Stage_Partial_Network_CVPR_2021_paper.html

• 1770 FPS - on GPU RTX 2080Ti - (416x416, fp16, batch=4) tkDNN/TensorRT https://github.com/ceccocats/tkDNN/issues/59#issuecomment-652269964

• 1353 FPS - on GPU RTX 2080Ti - (416x416, fp16, batch=4) OpenCV (including: transfering CPU->GPU and GPU->CPU) (excluding: nms, pre/post-processing) https://github.com/AlexeyAB/darknet/issues/6067#issuecomment-656604015

• 39 FPS - 25ms latency - on Jetson Nano - (416x416, fp16, batch=1) tkDNN/TensorRT https://github.com/ceccocats/tkDNN/issues/59#issuecomment-652157334

• 290 FPS - 3.5ms latency - on Jetson AGX - (416x416, fp16, batch=1) tkDNN/TensorRT https://github.com/ceccocats/tkDNN/issues/59#issuecomment-652157334

• 20 FPS on CPU ARM Kirin 990 - Smartphone Huawei P40 https://github.com/AlexeyAB/darknet/issues/6091#issuecomment-651502121 - Tencent/NCNN library https://github.com/Tencent/ncnn

• 120 FPS on nVidia Jetson AGX Xavier - MAX_N - Darknet framework

• 371 FPS on GPU GTX 1080 Ti - Darknet framework

---

• repository: https://github.com/AlexeyAB/darknet

• cfg: https://raw.githubusercontent.com/AlexeyAB/darknet/master/cfg/yolov4-tiny.cfg

• weights: https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-tiny.weights

• discussion: https://github.com/AlexeyAB/darknet/issues/6067

---

[AlexeyAB]

Discussion: https://www.reddit.com/r/MachineLearning/comments/hu7lyt/p_yolov4tiny_speed_1770_fps_tensorrtbatch4/

[kgksl]

Thanks for the excellent work.

My question is, is there a reason for tiny models to have a lower mAP when we set height and width of network to large values such as 608? When I performed inference with YOLOv4-tiny pertained models with COCO 2017 Validation set , I got the following results:

For 416x416 AP @[IoU=0.50:0.95] = 0.221 AP @ [IoU=0.50] = 0.406

For 608x608 AP @[IoU=0.50:0.95] = 0.187 AP @ [IoU=0.50] = 0.368

[AlexeyAB]

YOLOv4-tiny was trained for 416x416 (we used random=0, so it was trained without random shapes). You need to re-train YOLOv4-tiny with 608x608.

[kgksl]

YOLOv4-tiny was trained for 416x416 (we used random=0, so it was trained without random shapes). You need to re-train YOLOv4-tiny with 608x608.

Okay. Thanks a lot for the quick reply.