2020-12-04

App Improvements

Color Crop Embeddings by Image Metadata

Now you can color crops in the visualizer by image metadata to have even more ways to find patterns/errors in your dataset!

Move Elements Between Issues

A new button allows you to move elements from one issue to another

Fullscreen View

Fullscreen buttons have been added to main "Explore" page as well as the Issue Viewer so that the element grid can be expanded to allow for ease of use on small screens.

Editable Issue States

Top-level issue states are now editable! You may have seen this column on the main "Issues" page:

You can edit issue states by clicking on a specific issue (to go to the detailed Issue view), and clicking on the colored "Issue State" dropdown.

Updating issue states can help you keep track of your workflow.

Issue Element Downloads: Confusion Metadata!

Now, when you download crop-based issue elements, we will include metadata about what ground truth (if any) those crop detections were associated with.

Note that this is only available for new issue elements (not existing ones) added in one of two ways:

(1) When comparing a GT dataset and inference dataset > using the Embedding view > using the Crops tab > selecting the coloring scheme based on “confusion"

(2) When comparing a GT dataset and inference dataset > using the Confusion view > clicking on one of the cells in the confusion matrix (which should list a bunch of crops below)

Downloading as JSON: The format is an array of issue element objects. Each issue element’s metadata should be available as an label_metadata top-level field and look something like the following:

"label_metadata": {
      "confidence": 1.3357739448547363,
      "confidenceThreshold": 0, // the query setting when the issue was added
      "gtLabelClassId": 0,
      "gtLabelId": "006972_1_gt",
      "infLabelClassId": 3,
      "infLabelId": "006972_0_1_w_noisepred",
      "iou": 0.7369649410247803,
      "iouThreshold": 0.5 // the query setting when the issue was added
    },

Downloading as CSV: The fields above will not be nested, and will exist as columns per issue element (row).

Additional note: For false positives/negatives, the corresponding label IDs will be "", their label class IDs will be -1, and their iou will be 0. The confidence will be -1 for a false negative (no prediction).

Python Client / Data Upload Improvements

Upload Progress Tracking and Improvements

Uploading data in the newest versions of the client support rich tracking information of the upload and processing progress. Further, improvements have been made to the uploader to make it more resilient to transient failures.

Error Messages on Processing

The Aquarium Client and App now show error messages relating to the nature of failure when processing datasets and inferences. Error messages in the client show as follows

Dataset processing has failed. Exiting...
    generate_tsne: Attempted to layout embeddings multiple times. 
    Please contact Aquarium if you see this.

And in the Projects view in the App, when you hover over the colored dataset as follows:

Preview Frame

The client now supports the ability to generate preview frames before dataset upload. By creating a Projects (with filled out Label Class Maps) and passing a frame into the new method, we can create one time preview URLs so you can validate that your frame data and labels work as expected.

>>> al_client.preview_frame(project_id="demo", labeled_frame=labeled_frame)

Please visit the following url to preview your frame in the webapp
https://illume.aquariumlearning.com/projects/demo/preview_frame/e7402545f94b425a9255dad1704fa562
Please vist above URL to see your Preview frame.

You can also access a preview frame from the first frame in a dataset using the create_dataset method in the client, via a new arg supported called preview_first_frame

>>> al_client.create_dataset( ... , preview_first_frame=True)

Creating Dataset

Please visit the following url to preview your frame in the webapp
https://illume.aquariumlearning.com/projects/demo/preview_frame/e7402545f94b425a9255dad1704fa562
Please vist above URL to see your Preview frame.

Press ENTER to continue or type `exit` and press ENTER to cancel dataset upload.

Geo Data (Preview!)

The newest client now supports uploading of geo data per dataset frame using the EPSG:4326 WGS84 via Latitude/Longitude.

>>> frame.add_geo_latlong_data(37.81939245823887, -122.47846387094987)

This feature is still in preview. All geo data uploaded will remain intact, but App visualization tools for geo data are still not available. Stay posted for more information on this feature.

Last updated