Features

• Release YOLOv6-Segmentation models at full scales
• Achieve the state-of-the-art accuracy in Real-time Instance Segmentation.

Performance of YOLOv6-seg models

Model	Size	mAPbox 50-95	mAPmask 50-95	SpeedT4 trt fp16 b1 (fps)	Params (M)	FLOPs (G)
YOLOv6-N	640	35.3	31.2	645	4.9	7.0
YOLOv6-S	640	44.0	38.0	292	19.6	27.7
YOLOv6-M	640	48.2	41.3	148	37.1	54.3
YOLOv6-L	640	51.1	43.7	93	63.6	95.5
YOLOv6-X	640	52.2	44.8	47	119.1	175.5

Table Notes

• All checkpoints are trained from scratch on COCO for 300 epochs without distillation.
• Results of the mAP and speed are evaluated on COCO val2017 dataset with the input resolution of 640×640.
• Speed is tested with TensorRT 8.5 on T4 without post-processing.

v4.0 release

Features

• Release YOLOv6Lite models on mobile or CPU.
• Update MBLABlock in the network structure.
• Update YOLOv6Lite-face models on mobile or CPU.
• Code reconstruction and normalization of convolution operators.

Performance of YOLOv6Lite models

Model	Size	mAPval 0.5:0.95	sm8350 (ms)	mt6853 (ms)	sdm660 (ms)	Params (M)	FLOPs (G)
YOLOv6Lite-S	320*320	22.4	7.99	11.99	41.86	0.55	0.56
YOLOv6Lite-M	320*320	25.1	9.08	13.27	47.95	0.79	0.67
YOLOv6Lite-L	320*320	28.0	11.37	16.20	61.40	1.09	0.87
YOLOv6Lite-L	320*192	25.0	7.02	9.66	36.13	1.09	0.52
YOLOv6Lite-L	224*128	18.9	3.63	4.99	17.76	1.09	0.24

Table Notes

• From the perspective of model size and input image ratio, we have built a series of models on the mobile terminal to facilitate flexible applications in different scenarios.
• All checkpoints are trained with 400 epochs without distillation.
• Results of the mAP and speed are evaluated on COCO val2017 dataset, and the input resolution is the Size in the table.
• Speed is tested on MNN 2.3.0 AArch64 with 2 threads by arm82 acceleration. The inference warm-up is performed 10 times, and the cycle is performed 100 times.
• Qualcomm 888(sm8350), Dimensity 720(mt6853) and Qualcomm 660(sdm660) correspond to chips with different performances at the high, middle and low end respectively, which can be used as a reference for model capabilities under different chips.
• Refer to Test NCNN Speed tutorial to reproduce the NCNN speed results of YOLOv6Lite.

Performance of YOLOv6_MBLA models

Model	Size	mAPval 0.5:0.95	SpeedT4 trt fp16 b1 (fps)	SpeedT4 trt fp16 b32 (fps)	Params (M)	FLOPs (G)
YOLOv6-S-mbla	640	47.0distill	300	424	11.6	29.8
YOLOv6-M-mbla	640	50.3distill	168	216	26.1	66.7
YOLOv6-L-mbla	640	52.0distill	129	154	46.3	118.2
YOLOv6-X-mbla	640	53.5distill	78	94	78.8	199.0

Table Notes

• Speed is tested with TensorRT 8.4.2.4 on T4.
• The processes of model training, evaluation, and inference are the same as the original ones. For details, please refer to this README.

Features

• Face detection and landmarks localization
• Repulsion loss
• Same-channel Dehead

Performance on WIDERFACE

Model	Size	Easy	Medium	Hard	SpeedT4 trt fp16 b1 (fps)	SpeedT4 trt fp16 b32 (fps)	Params (M)	FLOPs (G)
YOLOv6-N	640	95.0	92.4	80.4	797	1313	4.63	11.35
YOLOv6-S	640	96.2	94.7	85.1	339	484	12.41	32.45
YOLOv6-M	640	97.0	95.3	86.3	188	240	24.85	70.59
YOLOv6-L	640	97.2	95.9	87.5	102	121	56.77	159.24

• All checkpoints are fine-tuned from COCO pretrained model for 300 epochs without distillation.
• Results of the mAP and speed are evaluated on WIDER FACE dataset with the input resolution of 640×640.
• Speed is tested with TensorRT 8.2 on T4.
• Refer to Test speed tutorial to reproduce the speed results of YOLOv6.
• Params and FLOPs of YOLOv6 are estimated on deployed models.

v3.0 release

Features

Release P6 models and update P5 models

• Renew the neck of the detector with a BiC module and SimCSPSPPF Block.
• Propose an anchor-aided training (AAT) strategy.
• Involve a new self-distillation strategy for small models of YOLOv6.
• Expand YOLOv6 and hit a new SOTA performance on the COCO dataset.

Performance

Model	Size	mAPval 0.5:0.95	SpeedT4 trt fp16 b1 (fps)	SpeedT4 trt fp16 b32 (fps)	Params (M)	FLOPs (G)
YOLOv6-N	640	37.5	779	1187	4.7	11.4
YOLOv6-S	640	45.0	339	484	18.5	45.3
YOLOv6-M	640	50.0	175	226	34.9	85.8
YOLOv6-L	640	52.8	98	116	59.6	150.7
YOLOv6-N6	1280	44.9	228	281	10.4	49.8
YOLOv6-S6	1280	50.3	98	108	41.4	198.0
YOLOv6-M6	1280	55.2	47	55	79.6	379.5
YOLOv6-L6	1280	57.2	26	29	140.4	673.4

Performance of base models

Model	Size	mAPval 0.5:0.95	SpeedT4 TRT FP16 b1 (FPS)	SpeedT4 TRT FP16 b32 (FPS)	SpeedT4 TRT INT8 b1 (FPS)	SpeedT4 TRT INT8 b32 (FPS)	Params (M)	FLOPs (G)
YOLOv6-N-base	640	36.6	727	1302	814	1805	4.65	11.46
YOLOv6-S-base	640	45.3	346	525	487	908	13.14	30.6
YOLOv6-M-base	640	49.4	179	245	284	439	28.33	72.30
YOLOv6-L-base	640	51.1	116	157	196	288	59.61	150.89

v2.1 release

Features

Release base models

• Use only regular convolution and Relu activation functions.

• Apply CSP (1/2 channel dim) blocks in the network structure, except for Nano base model.

Advantage:

• Adopt a unified network structure and configuration, and the accuracy loss of the PTQ 8-bit quantization model is negligible, about 0.4%.
• Suitable for users who are just getting started or who need to apply, optimize and deploy an 8-bit quantization model quickly and frequently.

Shortcoming:

• The accuracy on COCO is slightly lower than the v2.0 released models.

Performance

Model	Size	mAPval 0.5:0.95	SpeedT4 trt fp16 b1 (fps)	SpeedT4 trt fp16 b32 (fps)	Params (M)	FLOPs (G)
YOLOv6-N-base	640	35.6400e	832	1249	4.3	11.1
YOLOv6-S-base	640	43.8400e	373	531	11.5	27.6
YOLOv6-M-base	640	48.8distill	179	246	27.7	68.4
YOLOv6-L-base	640	51.0distill	115	153	58.5	144.0

v2.0 release

YOLOv6 has a series of models for various industrial scenarios, including nano/tiny/s/m/l, which the architectures vary considering the model size for better accuracy-speed trade-off. And some Bag-of-freebies methods are introduced to further improve the performance, such as self-distillation and more training epochs. For industrial deployment, we adopt QAT with channel-wise distillation and graph optimization to pursue extreme performance.

New Features

• Release M/L models and update N/T/S models with enhanced performance.⭐️ Benchmark
• 2x faster training time.
• Fix the degration of performance when evaluating on 640x640 inputs.
• Customized quantization methods. 🚀 Quantization Tutorial

v1.0 release

Features

YOLOv6 is a single-stage object detection framework dedicated to industrial application, with hardware-friendly efficient design and high performance, outperforming YOLOv5, YOLOX and PP-YOLOE.

YOLOv6-nano achieves 35.0 mAP on COCO val2017 dataset with 1242 FPS on T4 using TensorRT FP16 for bs32 inference, and YOLOv6-s achieves 43.1 mAP on COCO val2017 dataset with 520 FPS on T4 using TensorRT FP16 for bs32 inference.

• Hardware-friendly Design for Backbone and Neck
• Efficient Decoupled Head with SIoU Loss