How does YOLOX compare to YOLOR?

[augmentedstartups]

How does YOLOX compare to YOLOR by Wong Kin Yiu? In terms of speed and accuracy.

[GOATmessi8]

For yolo-r, I couldn't find a clear model for a fair comparison. Maybe the following model to compare with YOLOX-L (50.0% AP)?

Model	Test Size	APval	AP50val	AP75val	APSval	APMval	APLval	batch1 throughput
YOLOv4-CSP	640	49.1%	67.7%	53.8%	32.1%	54.4%	63.2%	76 fps
YOLOR-CSP	640	49.2%	67.6%	53.7%	32.9%	54.4%	63.0%	-

But I also notice that its training scripts seem to train 1200 epochs (300+450+450) as:

python -m torch.distributed.launch --nproc_per_node 8 --master_port 9527 train.py --batch-size 64 --img 1280 1280 --data data/coco.yaml --cfg cfg/yolor_p6.cfg --weights '' --device 0,1,2,3,4,5,6,7 --sync-bn --name yolor_p6 --hyp hyp.scratch.1280.yaml --epochs 300
python -m torch.distributed.launch --nproc_per_node 8 --master_port 9527 tune.py --batch-size 64 --img 1280 1280 --data data/coco.yaml --cfg cfg/yolor_p6.cfg --weights 'runs/train/yolor_p6/weights/last_298.pt' --device 0,1,2,3,4,5,6,7 --sync-bn --name yolor_p6-tune --hyp hyp.finetune.1280.yaml --epochs 450
python -m torch.distributed.launch --nproc_per_node 8 --master_port 9527 train.py --batch-size 64 --img 1280 1280 --data data/coco.yaml --cfg cfg/yolor_p6.cfg --weights 'runs/train/yolor_p6-tune/weights/epoch_424.pt' --device 0,1,2,3,4,5,6,7 --sync-bn --name yolor_p6-fine --hyp hyp.finetune.1280.yaml --epochs 450

while yolox models are trained with 300 epochs.

For larger scale model (1280x1280 input), we didn't cover it yet. And yolo-r did a good work.

[WongKinYiu]

Hello,

Model	Test Size	APval	AP50val	AP75val	APSval	APMval	APLval	batch1 throughput
**YOLOR-CSP***	640	50.0%	68.7%	54.3%	34.2%	55.1%	64.3%	76 fps
**YOLOR-CSP-X***	640	51.5%	69.9%	56.1%	35.8%	56.8%	66.1%	53 fps

We convert these two models into darknet weights, and batch1 throughput includes pre/post-processing. You could find them in https://github.com/AlexeyAB/darknet#pre-trained-models , which are yolov4-csp-swish and yolov4-csp-x-swish.

By the way, we follow the training schedule of Scaled-YOLOv4, so YOLOR-P5 models are trained 300 epochs. The training scripts of YOLOR-P6 models use resume training, so they are totally trained 450 epochs. We will release our new training script of YOLOR-P6 models which get about 0.6% better AP with 300 epochs training.

[GOATmessi8]

@WongKinYiu Great! And we hope some techniques in our YOLOX may be helpful for YOLOR too!

[WongKinYiu]

Yes, I have notice your OTA work for a long time. In these days, I have finish implemented decoupled head, anchor-free, and multi positives in our PyTorch version. And keep try hard to implement dynamic top-k assignment.

[fanfan-hanyang]

Can you explain the advantages and disadvantages of YOLOX and YOLOR when comparing？

[augmentedstartups]

Maybe combine the 2 to get YOLO XR or YOLO RX 🤣🤣

[augmentedstartups]

From Papers with Code it seems that YOLOR is better in mAP - https://paperswithcode.com/sota/object-detection-on-coco?tag_filter=15

[WongKinYiu]

+ Transformer = RTX

https://github.com/Megvii-BaseDetection/YOLOX#comming-soon

[augmentedstartups]

• Transformer = RTX

https://github.com/Megvii-BaseDetection/YOLOX#comming-soon

🤣🤣

[augmentedstartups]

This may be a bit unrelated but Iv posted a demo of YOLOR + DeepSORT Tracking https://youtu.be/keXpp8FhORI

[WongKinYiu]

Currently our YOLOR supports classification/detection/segmentation/embedding/tracking/reconstruction/landmark_detection in one unified model.

[augmentedstartups]

I see the example to run detection. How do I run it for segmentation, classification, etc?

[WongKinYiu]

@ruinmessi Hello, After apply simOTA on yolor-csp, it finally increase about 1% AP on mscoco object detection task. Thanks for your great work.

Train from scratch for 300 epochs:	Model	SimOTA	Test Size	APval
YOLOR-CSP	no	640	50.0%
YOLOR-CSP	yes	640	51.0%