I would like to embed predictions back into a web application for other users to consume, so that way we can upload large pools of information and make the results presented back to different departments categorised on risk scores and other correlating factors.

Currently if you use long tags, they will not be fully visible unless you hover on them. Enhance the UI so that tags are not cut halfway, and make it responsive.

Is there a way to reduce false positives even when there are few negative images that can be prepared?

Depending on what you want Custom Vision to learn, you may not be able to prepare a sufficient amount of images and variations. I would like to know if there are other things that can be done in this case.

When selecting & tagging the regions in my Workspace Images, there are often regions that are diagonal (not rotated at right angles). Selecting these regions will obviously cause the regions to include other parts of the image, causing the training to be less accurate or correct, and require more Workspace Images before being appropriate for use. This is a very common scenario for me (and I am guessing many others), since we do not always hold our cameras perfectly vertical when taking pictures. Thank you.

It would be great to have a tensorflow js example of classification and object detection in the same ZIP file of the model

Thanks in Advance

In image classification, we can get the reason why this model classified as dog or cat, etc, if we introduce grad-cam mechanism as attached.
Can we have this feature in image classification moidule?

https://arxiv.org/pdf/1610.02391.pdf

The docs suggest creating a fake tag for images with nothing in them. This is going to take space in the model and try for it to optimize something which doesn't make sense. When I get an image with a false positive I want to be able to upload the image to the training set even though there is nothing of interest in the picture.

The general domain results in 98% mAP, while the compact domain is 75% (for my data set). I am using a 4K camera to detect objects in real time and would like to be able to download and work with the general model.

Ability to simply backup or even copy a project.
The delete button is easily available, please add a copy button/feature.
Thank you,
Jared

The current bounding box method can't always be drawn over the exact area of the desired object.

A paint selection segmentation map or line tracing might speed up the laborious tagging process.

Can you get details about the Algorithms which is being used to train custom vision model ? Also details about how many layers of training was done.

Have the possibility to cancel a training.
I setted up an iteration to training during 6 hours, and after 12 hours the training continues and I can't do anything.

We're working on a model to recognize clothe types for an app. We've tested both object detection and classification, and classification is about 6.5% more accurate, most likely because of the specialized Retail domain. However, this domain is not available for object detection.

Can you add the pre-built retail domain to object detection projects?

While training data, people use the right hand to draw boxes. Buttons near the left hand, tab, w-a-s-d need to be utilized for fast labelling. a-d for image right/left, w-s for up and down through labels.
It makes labelling much much faster.

It would be great if we could export our customvision ML directly into the Azure Container Registry

When I click the tick in the bottom left hand corner of the training image the image just shrinks briefly and the tick doesn't turn blue. To be able to get the tick to turn blue which means it is selected, I need to hold down the left mouse key when I click on the tick and then move the cursor back over the tick until it turns light gray. This makes selecting multiple images very time consuming and fiddly. I've tested on Chrome, Firefox and Edge and all 3 browsers have the same problem.

provide confusion matrices in model evaluation