フィルターのクリア

training Object Detector AI with multiple ground truth objects, where each gTruth object might have a different datasource.

2 ビュー (過去 30 日間)
Greetings,
I have a question regarding Ground Truth objects.
Context:
I have many separate videos.
For each video, I made several ground truth objects using the video labeler app.
So, the 'source' property of each ground truth object might be different data source.
Also, each ground truth object might have different label definitions, depending on which objects appeared in that specific video.
Question:
I would like to create the object detector training data using all these different gTruth objects, where these gTruth objects have different sources, and also might have different label definitions.
Then, I will want to train the object detector with all the ground truth data from all those separate ground truth objects.
How would I do that?
Thank you,
William Murphy

回答 (1 件)

T.Nikhil kumar
T.Nikhil kumar 2024 年 4 月 10 日
Hello William,
I understand that you are trying to use ground truth data from different sources and different labels to train an object detector network.
This would require combining the ground truth data, ensuring consistent label definitions across all datasets, and then preparing the data for training. I would suggest you to try the following steps to achieve this:
  • Convert each of the ground truth objects into corresponding imageDatastore and boxLabelDatastore using the ‘objectDetectorTrainingData’ function. Combine these datastores to form a single combined datastore for that gTruth object. Then, combine all such datastores into a master combinedDatastore. You can use this combined datastore directly for training object detection models.
[imds1, blds1] = objectDetectorTrainingData(gTruth1);
[imds2, blds2] = objectDetectorTrainingData(gTruth2);
combinedDs1 = combine(imds1, blds1);
combinedDs2 = combine(imds2, blds2);
combinedDs = combine(combinedDs1, combinedDs2);
  • Ensure that all your gTruth objects use a consistent set of label definitions. If different videos have different objects and hence different labels, create a master list of all unique labels. You can manually go through the label definitions in each gTruth object and create the master list. You would require this list while creating a detector network.
  • You can then split your data for training and testing. You can also consider data augmentation to increase the diversity of the data.
  • Choose an appropriate detector network, configure training options, and finally train the object detector.
Refer to the following documentation to understand more about some functions used above:
Hope this helps!

カテゴリ

Help Center および File ExchangeImage and Video Ground Truth Labeling についてさらに検索

製品


リリース

R2022b

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by