Charts

Charts let you build custom aggregate visualizations over the Evaluation of an Experiment, alongside the Query Language and Embeddings exploration modes of the Evaluations interface.

Browsing Evaluation one image at a time is how you catch an individual failure; Charts are how you catch a pattern across hundreds or thousands of them. For a Computer Vision team, this kind of aggregate view is what turns "the model is sometimes wrong" into an actionable finding: a breakdown of average precision by DataTag can reveal that a specific acquisition condition (night, occlusion, a particular camera) is dragging down performance, and a breakdown by Label can reveal that one class is systematically confused with another.

1. Accessing Charts

Charts are available in two places:

• The Charts tab at the top of the Experiment, listing every chart created so far.

• The Charts option of the Evaluations exploration mode selector, which displays the Evaluation grid alongside the Charts panel for side-by-side exploration.

Clicking on a bar of a chart automatically filters the Evaluation displayed to only the ones matching that bar, by applying the equivalent query in the search bar, the same way it does on the Datalake and a DatasetVersion, detailed here. This drill-down is especially useful in the side-by-side exploration mode: spotting an anomaly on a chart and immediately visualizing the corresponding Evaluation, right next to it, ground truth and prediction included.

2. Creating a Chart

Click + Create chart to open the creation form:

• Name: the name of your chart.
• Aggregate Function: AVG, COUNT, MAX, MIN, or SUM.
• Aggregate Field: the field to aggregate — ID (to count Evaluation), CUSTOM_METRICS, or DATA_CUSTOM_METADATA.
• Group By Field: the field used to bucket the results. This field is optional: leaving it empty produces a Metric chart, a single aggregate value computed over the whole selection instead of a breakdown. You can group by:
  • The Evaluation's own COCO-style detection metrics, computed automatically by Picsellia for every Evaluation: AP_50_95, AP_50, AP_75, AP_50_95_SMALL, AP_50_95_MEDIUM, AP_50_95_LARGE, AR_50_95_DET1, AR_50_95_DET10, AR_50_95_DET100, AR_50_95_SMALL, AR_50_95_MEDIUM, AR_50_95_LARGE, TRUE_POSITIVES, FALSE_POSITIVES, and FALSE_NEGATIVES.
  • CUSTOM_METRICS, the metrics you logged yourself on the Evaluation (see the callout below).
  • Fields inherited from the underlying Data: DATA_CREATED_AT, DATA_CONTENT_TYPE, DATA_UPLOAD_STATUS, DATA_FILE_SIZE, DATA_TAG, DATA_SOURCE, DATA_METADATA, or DATA_CUSTOM_METADATA.
  • Fields inherited from the underlying Asset: ASSET_CREATED_AT or ASSET_TAG.
• Advanced filtering: optionally scope the chart to a subset of Evaluation using the same Query Language as the Search Bar.

For instance, a chart of the average AP_50_95 grouped by DATA_TAG lets you compare your ModelVersion performance across the different conditions tagged in your Data.

Once created and computed, your chart is displayed as a card showing the resulting visualization.

3. Managing Charts

From a chart's card, you can:

• Recompute: refresh the chart with the latest Evaluation.
• Update or Delete: from the ... menu.