Regarding, labeling ocluded objects. What if the object of interest is ocluded by another object that is not in our interest? Do we still label our object as we could see them in their entirety?
you should use a False Positive class in each model. When you want to use that model to auto annotate it will most likely generate false positive bounding boxes if used on images different from the ones from the dataset unless your model was trained on billions of images which is most unlikely, when the model thinks it identified a Positive but in fact is a False then you should change its class to False Positive and train the model with that. A model needs bad examples as well not only perfect examples, a model needs to be as precises as possible even when a class is partially occluded by another, for instance, in your case the Knight wasn't entirely in the bounding box at min 4:36 which i agree if the reason is not occlusion but if the object has enough details visible even when partially occluded, then it should be sampled as well. I'm enforcing what you said on 6:40
would u also recommend labelImg? its the annotating tool that the tutorial im watching is using, i dont really know to connect the roboflow to my files and the codes are pre-written
It was stated that all pawns need to be labeled so as to not confuse the model... but doesn't the training step only use the parts of the images that have the bounding boxes around it? IOW, if you label 4 pawns, but not say, 4 other pawns, does the training step even know those 4 other un-labeled pawns even exist?
Thank you man for helping, but i have question, if i have bounding boxes (xmin, xmax, ymin, ymax) with labels for all images in a csv file, how can i make it automatic in roboflow? because it will take a lot of time if i do this process manually.
didnt it occur to you that you can just a make a python script to convert the csv labels into text files and then use roboflow? i mean like how didnt occured to you? u an engineer?
i really need your help i couldnt make a tight bound box around objects my objects is so complicated and its 13 class , about 7000 images can you help me
thanks for such a nice video. what about occluded objects that are not shown in the picture and only some part of it is shown? Example would be three cars in a parking lot, and only 20% of third car is showing? should we include the part of car not showing or just that part which is showing?
create a labeled dataset of images by hand, say maybe 400, and then train a neural network using with your labeled dataset as input. The resulting model can then be used to predict the other 4600 images, and all you need to do is go through each prediction and verify the results
Hi Team Roboflow. I have a requirement, where I need to label images for fire/smoke use case. I need labeling for indoor as well as outdoor fires. Could you guys help me to produce those labels. Thanks
Thank you for grate tips. In my personal experience in some complex scenarios labeling occluded boxes as we can see it entirely would degrade model performance, especially if it is blocked by an object with different texture. What do you recommend in such a scenario.
@@Roboflow From my experience, you'll want an xml file for that image but you wont want a bounding box. This lets you show the model background images while not introducing a Null class.
Hello. It would be helpful if someone could give some advice about a problem I am facing. I am annotating an image of a road scene containing vehicles and there are vehicles present that are quite far away. Should I be annotating those as well?
Hi sir, I'm currently working on my university project and the project is about counting melon fruits from the tree. by far, I have 200 images with me. As of my progress now, i still did not managed to get the accurate results on the bounding box when i tested it after traning. One of the reasons could be because of how i labelled it. I find difficulty when labelling it because the melons on the tree are messy as they are surrounded with many leaves. so, most of the fruits are behind the leaves. So, my problem now is I'm not sure whether I'm labelling it the correct way and I have no one who is expertised regarding labelling to refer to at this moment. I hope you can at least give me your email so that i can approach you and get your guidance. I really really need some guidance from the right person. Hope you could recognise this comment sir. Thank you.
Hii nurul, I'm Milind this side, working on a similar fruit detection model as my master's thesis project where I'm experiencing the very scenario that u have mentioned here like occlusions due to leaves and overlapping fruits as well, if by now u have figured out in dealing with such cases, i would like to seek some guidance from you, if you don't mind sharing ur mail or any other details where we can connect...!
@@milindchaudhari1676 You should use models that are good at detecting small items. Also you should provide large, high quality images to the model versus smaller lower res images as that will help the model train better.
Another tip is to zoom in when labeling. The roboflow web labeled gives you a nice zoom tool so you can get tight bounding boxes.
Great tip! Thanks, Brian
When I zoom in I see a small reflection of the image..do I add that when bounding?
Great thanks alot
Regarding, labeling ocluded objects. What if the object of interest is ocluded by another object that is not in our interest? Do we still label our object as we could see them in their entirety?
you should use a False Positive class in each model. When you want to use that model to auto annotate it will most likely generate false positive bounding boxes if used on images different from the ones from the dataset unless your model was trained on billions of images which is most unlikely, when the model thinks it identified a Positive but in fact is a False then you should change its class to False Positive and train the model with that. A model needs bad examples as well not only perfect examples, a model needs to be as precises as possible even when a class is partially occluded by another, for instance, in your case the Knight wasn't entirely in the bounding box at min 4:36 which i agree if the reason is not occlusion but if the object has enough details visible even when partially occluded, then it should be sampled as well. I'm enforcing what you said on 6:40
would u also recommend labelImg? its the annotating tool that the tutorial im watching is using, i dont really know to connect the roboflow to my files and the codes are pre-written
Which Tool you used ?
Thank you.
Thank you, this makes my work easy. Keep making videos like this.
Hi, Is any circle Annotation for Object Detection . Present we are using square, polygon like that Any Circle Annotation for object detection.
It was stated that all pawns need to be labeled so as to not confuse the model... but doesn't the training step only use the parts of the images that have the bounding boxes around it? IOW, if you label 4 pawns, but not say, 4 other pawns, does the training step even know those 4 other un-labeled pawns even exist?
How to annotate football field? We can use polygon tool, but if you want to stop one corner and start from another corner, then we cannot. Please help
Thank you man for helping, but i have question, if i have bounding boxes (xmin, xmax, ymin, ymax) with labels for all images in a csv file, how can i make it automatic in roboflow? because it will take a lot of time if i do this process manually.
didnt it occur to you that you can just a make a python script to convert the csv labels into text files and then use roboflow? i mean like how didnt occured to you? u an engineer?
i really need your help i couldnt make a tight bound box around objects my objects is so complicated and its 13 class , about 7000 images can you help me
thanks for such a nice video.
what about occluded objects that are not shown in the picture and only some part of it is shown? Example would be three cars in a parking lot, and only 20% of third car is showing? should we include the part of car not showing or just that part which is showing?
Wondering the same thing. Did you figure it out?
block pawn💀☠☠☠
i need a tip on how to easly lable 5000 images
create a labeled dataset of images by hand, say maybe 400, and then train a neural network using with your labeled dataset as input. The resulting model can then be used to predict the other 4600 images, and all you need to do is go through each prediction and verify the results
Hi Team Roboflow. I have a requirement, where I need to label images for fire/smoke use case. I need labeling for indoor as well as outdoor fires. Could you guys help me to produce those labels. Thanks
How to annotate empty images, where there are no objects?
Thank you for grate tips. In my personal experience in some complex scenarios labeling occluded boxes as we can see it entirely would degrade model performance, especially if it is blocked by an object with different texture. What do you recommend in such a scenario.
Sir, what to do when there are no objects of interest in image? should those be annotated?
Yes! Annotate them as "Null" this lets the model know the object is not in the image. It's good to have many Null images in your dataset
@@Roboflow thank you sir...
@@Roboflow From my experience, you'll want an xml file for that image but you wont want a bounding box. This lets you show the model background images while not introducing a Null class.
Rotate the board 90 degrees.
Thanks a bunch for the helpful tips.
Hello. It would be helpful if someone could give some advice about a problem I am facing. I am annotating an image of a road scene containing vehicles and there are vehicles present that are quite far away. Should I be annotating those as well?
If they are far enough to be blurry in the image, then don't bother to annotate them. If they are visible enough then annotate them.
6:40
too expensive
You can always use free tier:)
hi how do I delete the wrong labels I tried so many ways just couldnt find a delete label or edit labeling. Thank you.
Try to select it, then press del (delete) on the keyboard
I want to change color for each labels
Hi sir, I'm currently working on my university project and the project is about counting melon fruits from the tree. by far, I have 200 images with me. As of my progress now, i still did not managed to get the accurate results on the bounding box when i tested it after traning. One of the reasons could be because of how i labelled it. I find difficulty when labelling it because the melons on the tree are messy as they are surrounded with many leaves. so, most of the fruits are behind the leaves. So, my problem now is I'm not sure whether I'm labelling it the correct way and I have no one who is expertised regarding labelling to refer to at this moment. I hope you can at least give me your email so that i can approach you and get your guidance. I really really need some guidance from the right person. Hope you could recognise this comment sir. Thank you.
Hi there! Please post in our forum in our community members can help discuss.roboflow.com/
Hii nurul, I'm Milind this side, working on a similar fruit detection model as my master's thesis project where I'm experiencing the very scenario that u have mentioned here like occlusions due to leaves and overlapping fruits as well, if by now u have figured out in dealing with such cases, i would like to seek some guidance from you, if you don't mind sharing ur mail or any other details where we can connect...!
@@milindchaudhari1676 You should use models that are good at detecting small items. Also you should provide large, high quality images to the model versus smaller lower res images as that will help the model train better.