04Classifiers and decoders
Temporal convolutional networks
Dilated one-dimensional convolutions that read seconds of context for every frame. The standard learned sequence layer for dense per-frame labels.
- Sequence
- Learned
- Needs labels
- MIT
- Modality
- A sequence of per-frame features
- Output
- Class scores for every frame
- Evidence
- Established TCN literature; maintained PyTorch package
- Code and weights
- pytorch-tcn package; original TCN reference code
- Licence
- MIT
A temporal convolutional network stacks one-dimensional convolutions whose dilation doubles at each layer. The first layer sees neighbouring frames, the next skips one, the next three. After a handful of layers a single output frame depends on hundreds of inputs, with far fewer parameters than a recurrent network and none of its training troubles. Residual connections keep the stack trainable.
Two variants matter. A causal TCN sees only the past and suits live inference. A non-causal one sees both sides and is the right choice offline, where the frames after a touch are as informative as the frames before it.
pytorch-tcn packages the architecture with streaming inference and export support. The locuslab repository is the reference code from the work that established TCNs as a general sequence model.
Where it wins
- Receptive field by design. At 25 fps, covering four seconds is an architectural choice, not a hope.
- Parallel over time. Inference on a full half is one forward pass.
- Stable to train compared with recurrent networks.
- The same layer serves tracking features, set-encoder outputs and video embeddings.
What it costs
- Supervised. It needs dense per-frame labels in quantity.
- Context is not coherence. Outputs are still per-frame scores and still flicker at ambiguous moments.
- The receptive field is fixed. Long dead-ball stretches exceed it; short duels are blurred by it.
- Trained weights are tied to a frame rate and a feature definition.
How it plugs into possession
In the middle of the stack. A set encoder summarises each frame, the TCN adds time, a decoder enforces sequence logic. Borghesi’s possession model is exactly this, and MS-TCN is what happens when the refinement job is handed to more TCN stages.
For offline possession, use the non-causal form. Whether a touch was a reception or a deflection is decided by what the ball does next.
Our read
The default learned temporal layer. It buys context cheaply. Coherence still has to be decoded.