YOLOv7 | Instance Segmentation on Custom Dataset

This video will show you step by step implementation of Instance Segmentation using YOLOv7.

YoloV7 is new framework which can perform various computer vision tasks like Object Detection, instance segmentation, keypoints detection .

This is the only framework support YOLOv4 + Instance Segmentation in single stage style.

It is based on detectron2. Detectron2 is A PyTorch-based modular object detection library by Facebook.
Yolov7 supports the training of yolov6.

[benaamediajo]

Notes for anyone who is facing issues:
1. copy the register commands to the training py file as seen in the video
2. change number of workers to 4, also max iterations to 3000 no 190000

[amirhosinipur]

Thanks a lot for the detailed instruction. Some issues I ran into did resolve by reading the comments. I am sure that helps in future production.
Since the training is taking at least 2 days on my laptop with RTX 3080, if I want to take advantage of a cloud GPU, like 8 A100, then what parameters should I change to make it faster. With the same settings and even increasing the "IMS_PER_BATCH" in the configs/Base-YOLOv7.yaml file to 256, it does not change the time (before and after the change it stays 1 day and 23 Hrs on 8 A100 GPU with each have 40 GB, but only 5 GB of each was being used).

[estermoiseyev]

I have a question, what if I labeled my images by myself with the labelImg tool, and I don't have the json file in the train folder with the annotation details of thos images, is there a way to create this file?

[zahrahajalioghli]

Thank you for the practical tutorials.
I have the following questions:

Can we use the saved weights from YOLOv7 instance segmentation for a classification problem?
We have a binary classification problem with 500 images, one class having only 30 images and the rest belonging to the other class. Can we extract features using instance segmentation on the images with fewer samples and then use all the features for classification?

[zahrahajalioghli]

Thank you for the practical tutorials.🙏🙏🙏
I have the following questions:

Can we use the saved weights from YOLOv7 instance segmentation for a classification problem?
We have a binary classification problem with 500 images, one class having only 30 images and the rest belonging to the other class. Can we extract features using instance segmentation on the images with fewer samples and then use all the features for classification?

[coder]

Does this version of yolov7 do semantic segmentation?

[shwetabhat]

Great content ma'am, requesting you to kindly continue with the series as it greatly helps . 🙂

[anushuyaslk]

I have registered my coco json using the code shown but when I start to train the model it shows dataset not registered what should I do

[JanikSteier]

Thank you very much for your content, Aarohi! It helped me a lot.

I'm facing a little issue in the end of procedure: When I excute the last command of the video to test model the test image occures proberly with the bounding boxes and the instance segmentation masks. The bounding boxes are on the correct postion around the ballons, but the instance segementation masks occur always in the upper left of the image, moved away from the actual ballons additional they are smaller as the should be. The general shape and amount of the masks is correct.

Do you know where the problem is and how to fix it ?

[rishabhsheoran]

Hey Aarohi! can you please help me with the deployment of the yolov7 segmnetation model on triton?
When I hit the triton inference server, I get back following outputs:
name: output
tensor: float32[batch, anchors, Concatoutput_dim_2]

name: onnx::Slice_539
tensor: float32[Transposeonnx::Slice_539_dim_0, 3, Transposeonnx::Slice_539_dim_2, Transposeonnx::Slice_539_dim_3, 40]


name: onnx::Slice_693
tensor: float32[Transposeonnx::Slice_693_dim_0, 3, Transposeonnx::Slice_693_dim_2, Transposeonnx::Slice_693_dim_3, 40]

name: onnx::Slice_844
tensor: float32[Transposeonnx::Slice_844_dim_0, 3, Transposeonnx::Slice_844_dim_2, Transposeonnx::Slice_844_dim_3, 40]

name: 517
tensor: float32[Mul517_dim_0, 32, Mul517_dim_2, Mul517_dim_3]

output <class 'numpy.ndarray'> (1, 100800, 40)
onnx::Slice_539 <class 'numpy.ndarray'> (1, 3, 160, 160, 40)
onnx::Slice_693 <class 'numpy.ndarray'> (1, 3, 80, 80, 40)
onnx::Slice_844 <class 'numpy.ndarray'> (1, 3, 40, 40, 40)

From the above outputs how do i extract the bounding boxes and the masks?

[مثنىالربيه]

thank you so much for sharing, can YOLOv7 be used with a pre-trained network such as AlexNet or GoogleNet ?

[sumyatnoe]

Hello Aarohi! Firstly thank u very much for your explaination. Firstly I have a problem with yolomask.yaml file. In this file the batch size and parameter are too large, My GPU which is RTX a 6000 can't hadle the training process. May I know about did you change the parameter or how can I make the efficient training?

[vishavjeetsharma]

Hy mam, I need some help, the problem on which I am working can be solved using this approach, I need some guidance, problem ( count no of rice grain, find it's width, height etc, in order to do that i need to instance segment each rice grain ( in case of overlaping rice grains), I have rice grain dataset for different varieties of rice, can you guide me how can i create a model, and using that model how can i segment each rice Am am a beginer, and dont know much about this yolo instance segmentation, pls help me

[Aaryav]

Thanks for the detailed video. Is it possible to use java for instance segmentation?

[devavratpro]

Thanks for this great Video. I would like to ask how object detection can be implemented using a rotated bounding box? Any suggestions, please!!!

[beeseah]

Hi Aarohi your tutorials are excellent! I would like to ask - when converting to coco data format, why did you add 0.5 to both the x and the y? Wouldn't that differ from the original outline of the object?

[anilghodekar]

Hello Aarohi. Thanks for the great work you are doing. Can we use YOLO V7 for copy-move image forgery detection? Can you pls explain at high level how would be the methodology?

[machinelearninglearning]

when i run the my_demo file i got photo without mask or bounding box (the original photo ) how can i fix this problem, can you please share with me the my_demo file

[rsalazar]

Thank you very much for your videos, they are very good! One question, YoloV7 will continue to be very fast and precise, even if there are only a few (6) objects to detect ?

[rakeshkumarrout]

Hi mam this is really useful. Mam how to get the location (bbox) coordinates of the predicted images?? Kindly help mam. The last output images coordinates. Pxpypwph