Frequently Asked Questions

Uploaded images appear completely black in the annotation page

../../_images/invalid_images.png

When uploading images, DaoAI World checks and validates the data. Some images may fail to upload (as shown above). This might be due to corrupted images. You are advised to re-capture the images or try the following:

Attempt to repair the corrupted data

Use the XnView MP image tool, which you can download from the following link: XnView MP

Extract the files, find the XnView MP executable, and double-click to run it.

../../_images/xnviewmp_exe.png

Go to File -> Open to select the corrupted image.

../../_images/xnviewmp_open_file.png

Go to File -> Save As and save the image to a new location.

../../_images/xnviewmp_save_as.png

Upload the saved image again.

../../_images/xnviewmp_upload_success.png

Model inference fails when the object is at the edge

If you find that the model fails to match the object and produces an error when the object is at the edge:

../../_images/roi_fail.png

Check whether the ROI (Region of Interest) setting was applied during model training and if the object appears outside the ROI.

../../_images/roi_fail_setting.png

Since the ROI preprocessing trims background areas and keeps only the area of interest, when an object is outside the ROI, it is considered part of the background and gets trimmed, resulting in no prediction.

To resolve this issue, consider:

  1. Updating the ROI area by redefining the ROI to ensure it covers the area where the object may appear. Then, retrain the model.

  2. Removing the ROI preprocessing and retraining the model.

This will allow you to correctly predict objects at the image's edge.

../../_images/roi_fail_result.png

Model inference fails due to different image resolutions

If the model fails during inference, check whether the resolution used during inference matches the resolution of the training data.

In general, models can tolerate some changes in resolution. However, if an ROI preprocessing step is involved, the ROI will be based on the original image resolution.

If you use a much lower resolution image for inference than the one used during training, the entire image may be cropped by the ROI, leading to inference failure.

For example, if the training images have a width of 8000 pixels and the ROI preprocessing crops 1000 pixels off each side, using a 1000-pixel wide image for inference will result in the ROI cropping out the entire image, causing inference to fail.

../../_images/res_fail.JPG

Ensure you use images with the same resolution as the training set when testing and inferring.