YOLOv4 model zoo

YOLOv4 model zoo

We provide a collection of COCO pre-trained files of YOLOv4 series models: https://hackmd.io/NFj2NrmqTcefjc2l94KjpQ?view#Model-descriptions

COCO pre-trained files (512x512)

Models used anchors of 512x512 input resolution for training.

Model	Size	AP	AP50	AP75	APS	APM	APL	cfg	weights	score
YOLOv4	512	43.0	64.9	46.5	24.3	46.1	55.2	cfg	weights	score
YOLOv4-Leaky	512	42.4	64.5	46.0	23.9	45.6	54.2	cfg	weights	score
YOLOv4-SAM-Leaky	512	42.7	64.5	46.4	24.1	45.3	54.8	cfg	weights	score
YOLOv4-Mish	512	43.2	64.7	46.9	24.4	46.1	55.4	cfg	weights	score
YOLOv4-SAM-Mish	512	43.4	65.0	47.2	24.6	46.7	55.5	cfg	weights	score

Model	Size	AP	AP50	AP75	APS	APM	APL	cfg	weights	score
YOLOv4	608	43.5	65.7	47.3	26.7	46.7	53.3	cfg	weights	score
YOLOv4-Leaky	608	42.9	65.3	46.8	26.1	46.3	52.3	cfg	weights	score
YOLOv4-SAM-Leaky	608	43.3	65.4	47.1	26.2	46.1	53.2	cfg	weights	score
YOLOv4-Mish	608	43.8	65.6	47.8	26.7	46.8	53.6	cfg	weights	score
YOLOv4-SAM-Mish	608	44.0	65.7	48.0	26.7	47.2	54.1	cfg	weights	score

COCO pre-trained files (416x416)

Models used anchors of 416x416 input resolution for training.

Model	Size	AP	AP50	AP75	APS	APM	APL	cfg	weights	score
YOLOv4-Leaky-416	416	40.7	62.7	43.9	21.4	43.7	54.0	cfg	weights	score
YOLOv4-Mish-416	416	41.5	63.3	44.7	21.9	44.4	55.3	cfg	weights	score

Select which model is suitable for training on your GPU

Here we list the minimum requirement GPU-RAM for training different models. However, we do not suggest you train the model with subdivisions equal or larger than 32, it will takes very long training time. We are developing the training method that can use less GPU-RAM to train a model.

• for 4GB-RAM GPUs
  • YOLOv4-Leaky-416: batch=64, subdivisions=64.
• for 6GB-RAM GPUs
  • YOLOv4-Leaky-416: batch=64, subdivisions=32.
  • YOLOv4-Mish-416: batch=64, subdivisions=64.
• for 8GB-RAM GPUs
  • YOLOv4-Mish-416: batch=64, subdivisions=32.
• for 11GB-RAM GPUs (1080ti, 2080ti)
  • YOLOv4-Leaky-416: batch=64, subdivisions=16.
  • YOLOv4-Mish-416: batch=64, subdivisions=16.
• for 12GB-RAM GPUs (titan x, titan xp, titan v)
  • YOLOv4-Leaky: batch=64, subdivisions=16.
  • YOLOv4-SAM-Leaky: batch=64, subdivisions=16.
• for 16GB-RAM GPUs (p100)
  • YOLOv4: batch=64, subdivisions=16.
  • YOLOv4-Mish: batch=64, subdivisions=16.
  • YOLOv4-SAM-Mish: batch=64, subdivisions=16.
• for 24GB-RAM GPUs (titan rtx, rtx 6000)
  • YOLOv4-Leaky: batch=64, subdivisions=8.
  • YOLOv4-SAM-Leaky: batch=64, subdivisions=8.
• for 32GB-RAM GPUs (v100, v100s)
  • YOLOv4: batch=64, subdivisions=8.
  • YOLOv4-Mish: batch=64, subdivisions=8.
  • YOLOv4-SAM-Mish: batch=64, subdivisions=8.

Model descriptions

• YOLOv4

  • The model in YOLOv4 paper.
  • Backbone - CSPDarknet53 with Mish activation
  • Neck - PANet with Leaky activation
  • Plugin Modules - SPP
  • V100 FPS - 62@608x608, 83@512x512
  • BFLOPs - 128.5@608x608

• YOLOv4-Leaky

  • Backbone - CSPDarknet53 with Leaky activation
  • Neck - PANet with Leaky activation
  • Plugin Modules - SPP
  • BFLOPs - 128.5@608x608

• YOLOv4-SAM-Leaky

  • Backbone - CSPDarknet53 with Leaky activation
  • Neck - PANet with Leaky activation
  • Plugin Modules - SPP, SAM
  • BFLOPs - 130.7@608x608

• YOLOv4-Mish

  • Backbone - CSPDarknet53 with Mish activation
  • Neck - PANet with Mish activation
  • Plugin Modules - SPP
  • BFLOPs - 128.5@608x608

• YOLOv4-SAM-Mish

  • Backbone - CSPDarknet53 with Mish activation
  • Neck - PANet with Mish activation
  • Plugin Modules - SPP, SAM
  • V100 FPS - 61@608x608, 81@512x512
  • BFLOPs - 130.7@608x608

• YOLOv4-CSP

  • will update
  • Backbone - CSPDarknet53 with Mish activation
  • Neck - CSPPANet with Mish activation
  • Plugin Modules - SPP

• YOLOv4-CSP-SAM

  • will update
  • Backbone - CSPDarknet53 with Mish activation
  • Neck - CSPPANet with Mish activation
  • Plugin Modules - SPP, SAM