Metrics¶
Tinify provides several metrics to evaluate the quality and performance of compression models. These metrics are commonly used in image and video compression research.
Image Metrics¶
Quality Metrics¶
PSNR (Peak Signal-to-Noise Ratio)¶
The Peak Signal-to-Noise Ratio (PSNR) is a widely used metric to measure the quality of reconstruction of lossy compression codecs. It is defined as the ratio between the maximum possible power of a signal and the power of corrupting noise that affects the fidelity of its representation.
- psnr-rgb: PSNR calculated on the RGB channels.
MS-SSIM (Multi-Scale Structural Similarity Index)¶
The Multi-Scale Structural Similarity Index (MS-SSIM) provides a quality measure that is often more consistent with human visual perception than PSNR. It is calculated over multiple scales.
- ms-ssim-rgb: MS-SSIM calculated on the RGB channels.
Compression Metrics¶
BPP (Bits Per Pixel)¶
Bits Per Pixel (bpp) measures the average number of bits used to store one pixel of the image. It is calculated as:
Performance Metrics¶
- encoding_time: Time taken to encode the image (in seconds).
- decoding_time: Time taken to decode the image (in seconds).
Video Metrics¶
Quality Metrics¶
Video quality is often evaluated on both RGB and YCbCr color spaces.
YCbCr Metrics¶
- psnr-y: PSNR calculated on the Luma (Y) channel.
- psnr-u: PSNR calculated on the Cb (U) channel.
- psnr-v: PSNR calculated on the Cr (V) channel.
- psnr-yuv: A weighted average of the Y, U, and V PSNRs, typically calculated as: $$ \text{PSNR}_{YUV} = \frac{4 \cdot \text{PSNR}_Y + \text{PSNR}_U + \text{PSNR}_V}{6} $$
RGB Metrics¶
- psnr-rgb: PSNR calculated on the RGB channels.
- ms-ssim-rgb: MS-SSIM calculated on the RGB channels.
- mse-rgb: Mean Squared Error (MSE) calculated on the RGB channels.
Compression Metrics¶
Bitrate¶
For video, compression efficiency is often measured in bitrate (e.g., kbps) or bits per pixel.
- bitrate: The average bitrate of the video stream.
- bpp: Bits Per Pixel, averaged over all frames.
Performance Metrics¶
- encoding_time: Total time taken to encode the video sequence.
- decoding_time: Total time taken to decode the video sequence.
Loss Metrics¶
During training, the following metrics are often used as components of the loss function:
- mse_loss: Mean Squared Error between the original and reconstructed images/frames.
- ms_ssim_loss: \(1 - \text{MS-SSIM}\).
- bpp_loss: The estimated bitrate (entropy) of the latent representation.