Video Codec — An Introduction

Video codecs are the backbone of modern media streaming, videoconferencing, and storage. Standards like H.264 [1] and H.265/HEVC [2] achieve impressive compression ratios by exploiting both spatial and temporal redundancy in video. This post walks through the core encoder–decoder pipeline shared by most traditional video codecs.

Encoder

Step 1: Motion Estimation

The encoder estimates the motion between the current frame $x_t$ and the previous reconstructed frame $x_{t-1}$ stored in the Decoded Picture Buffer (DPB). This produces a motion vector $v_t$ for each macro-block, describing how regions of the image have shifted between frames.

Step 2: Motion Compensation

Using $v_t$, the encoder warps the previous frame to produce a predicted image $\bar{x}_t$. The difference between the current and predicted frame — the residual — is then computed as:

Because consecutive frames in a video tend to look similar, $r_t$ is typically small and sparse, making it much cheaper to encode than the full frame.

Step 3: Transform and Quantization

The residual $r_t$ is transformed into the frequency domain using the Discrete Cosine Transform (DCT) [4]. DCT concentrates most of the signal energy into a small number of low-frequency coefficients, and since the human visual system is far less sensitive to high-frequency detail, those components can be aggressively discarded.

The transformed coefficients are then quantized to produce $\hat{y}_t$. Quantization introduces a controlled loss of precision — it is the primary source of compression loss in traditional codecs, and the quantization parameter (QP) directly controls the quality–bitrate trade-off. A key side benefit of DCT is that it reduces the variance of the data, which allows the subsequent entropy encoder to achieve better compression.

Step 4: Inverse Transform

To prepare the DPB for encoding the next frame, the encoder reconstructs an approximation of the current frame. The quantized coefficients $\hat{y}_t$ are passed through the inverse DCT to recover the reconstructed residual $\hat{r}_t$, which is then added back to the predicted image:

This reconstructed frame $\hat{x}_t$ is stored in the DPB. Note that this step mirrors exactly what the decoder will do, ensuring encoder and decoder stay in sync.

Step 5: Entropy Coding

Finally, the quantized residual $\hat{y}_t$ and the motion vectors $v_t$ are losslessly compressed using an entropy coder (e.g., CABAC in H.264/H.265) and packed into the output bitstream.

Decoder

The decoder is structurally simpler than the encoder — it runs a subset of the same operations.

Step 1: Entropy Decoding

The decoder parses the incoming bitstream to recover the quantized residual $\hat{y}_t$ and the motion vectors $v_t$.

Step 2: Inverse Transform

The residual $\hat{y}_t$ is passed through the inverse DCT to convert it back to pixel-domain residual $\hat{r}_t$.

Step 3: Motion Compensation

The decoder maintains its own DPB. Using the motion vectors $v_t$, it warps the previously decoded frame $\hat{x}_{t-1}$ to predict the current frame $\bar{x}_t$.

Step 4: Frame Reconstruction

The predicted frame $\bar{x}_t$ and the residual $\hat{r}_t$ are combined to reconstruct the final output frame:

This $\hat{x}_t$ is stored in the decoder’s DPB and displayed to the user.

References