Understand IOU Distribution
Get an overview of prediction tightness
With this workflow you can easily obtain distribution of IOU on matches to roughly understand how well the model is performing on tightness and to get to the visual samples of any range of interest fast.
Pre-reqs: Dataset, Annotations, Predictions (evaluated)
- Open dataset with annotations & predictions both uploaded
- Use the top navigation to go to the “Charts” page
- Use the drop down to select model of interest
- Scroll down to the “Match IOU Distribution” chart
- Hover on any bar to get an absolute number for any interval in the range
- See distribution over a specific subset by running a query e.g. metadata.weather=rain
- Click on any bar to query the underlying data items in the grid view
Ideally you would want this distribution to be right skewed i.e. the predictions are very tight. This is especially important in domains like self-driving where the degree of overlap of the prediction with annotation matters a lot.
Updated about 1 year ago