📎Image Metrics
The following image metrics are computed during the training and also during the inference and the drift in these metrics are monitored to understand the distribution of the data during the inference and training.
Image Brightness: This metric measures the overall intensity of light in an image. Monitoring brightness helps detect shifts in lighting conditions between training and inference data. A drift in brightness can indicate different environments, like changes in daytime or artificial lighting. By comparing brightness distribution during training and inference, you can ensure the model is adapting to real-world conditions.
Image Noise: Noise refers to random variations in pixel values that can obscure important details. Monitoring image noise during both training and inference helps identify issues with data quality, such as sensor malfunctions or environmental factors. An increase in noise during inference compared to training data can signal degraded sensor performance or unexpected environmental changes.
Image Sharpness: Sharpness reflects the clarity of details in an image, often influenced by the focus and resolution. Monitoring sharpness ensures that the image quality remains consistent between training and inference. Sharpness drift could indicate differences in camera calibration, focus, or motion blur, affecting the model’s ability to extract features accurately.
Image Channel Mean: This metric measures the average pixel intensity for each color channel (e.g., RGB channels). Monitoring the channel mean helps track shifts in color balance, which may occur due to varying lighting conditions or sensor adjustments. Drift in this metric between training and inference can impact how the model interprets features related to color.
Image Pixel Distribution: This metric tracks the overall distribution of pixel values across the entire image. Monitoring pixel distribution helps detect broader changes in image characteristics, such as contrast, saturation, and overall tone. Shifts in pixel distribution between training and inference indicate potential changes in the visual environment, affecting model performance.
Last updated