fcjian
what is the off-the-shelf detector used to label the synthetic images?
Can the detector that is trained with real + synthetic images achieve comparable performance to the off-the-shelf detector (or improve?) The logic here looks weird.
The off-the-shelf detectors are solely trained on images from a restricted set of base categories, resulting in their limited capacity to detect novel categories. In contrast, InstaGen has the capability to generate images across a wide range of categories and diversity, thus extending the scope of categories that the detector can recognize and enhancing its overall performance.
The off-the-shelf detectors are solely trained on images from a restricted set of base categories, resulting in their limited capacity to detect novel categories. In contrast, InstaGen has the capability to generate images across a wide range of categories and diversity, thus extending the scope of categories that the detector can recognize and enhancing its overall performance.
How do you create labels for the generated images in novel category? If you are using the off-the-shelf detector, isn't it not accurate for the novel category?
The off-the-shelf detectors are solely trained on images from a restricted set of base categories, resulting in their limited capacity to detect novel categories. In contrast, InstaGen has the capability to generate images across a wide range of categories and diversity, thus extending the scope of categories that the detector can recognize and enhancing its overall performance.
How do you create labels for the generated images in novel category? If you are using the off-the-shelf detector, isn't it not accurate for the novel category?
We trained a instance-level grounding head on the base categories, which can accurately predict bounding boxes for novel categories.