I have a training dataset which has many images but do not have bounding box annotations. I have annotations for the images in validation and test dataset. Goal is to train a detectron2 model using validation (with bbox) and training dataset(without bbox) without affecting the accuracy of the model. When I tried to train the model with validation dataset only the accuracy was 71% but when I tried to combine both train and validation dataset the accuracy reduced to 9%. I am using Faster RCNN is an object detection architecture (faster_rcnn_R_50_FPN_1x.yaml) Can someone please help me on how to train the model without the bounding box ?
Asked
Active
Viewed 642 times
0
-
I do not have much reputation to comment so I am writing here. Can you let us know what was tool which you used for annotations? Generally, when you are using any annotation tool and doing segmentations of the object you get BBox as well. Can you check your annotations whether there are any X and Y values? – yogi Apr 13 '21 at 08:12
-
Its not clear whether you want to improve your accuracy or train a custom dataset without bounding box annotations. If I am not wrong segmentation mask is an optional field. I don't think bbox is optional. – Aakash Gupta Aug 02 '21 at 15:32
-
Maybe sharing some sample data points will help in explaining the problem – Aakash Gupta Aug 02 '21 at 15:43