04Events and replay

E2E-Spot

End-to-end spotting of temporally precise events: per-frame features and a recurrent layer, trained together so that fine timing survives.

  • Video
  • Event spotting
  • Open weights
  • BSD-3-Clause
Every frame is encoded, context is added across time, and every frame gets a class or background.
Modality
Video
Output
Event class scores for every frame
Evidence
ECCV 2022
Code and weights
Training and inference code; models in a companion repository
Licence
BSD-3-Clause

“Spotting Temporally Precise, Fine-Grained Events in Video” named the task that ball-action challenges later adopted: find events at a tolerance of a frame or two, not a second or two. The paper’s argument is that this requires end-to-end training. Pipelines that freeze a feature extractor and train a head on top lose the fine temporal detail in the features.

E2E-Spot is the compact baseline that makes the case. A lightweight network encodes every frame, with temporal information mixed between neighbouring frames inside it. A recurrent layer aggregates over the clip. The model classifies every frame as an event class or background and trains directly on long clips.

Where it wins

  • Frame-level output by design. The task definition matches the precision possession needs.
  • Small and fast. Dense per-frame inference is affordable.
  • BSD-3-Clause, the most permissive licence among the spotters here.
  • A clean, general baseline that later football spotters build on and compare against.

What it costs

  • The released models target the paper’s benchmarks. Fine-grained ball actions on a new camera mean retraining.
  • Events are sparse and background is everywhere: extreme class imbalance, and sensitivity to timing noise in the labels.
  • No actor, no team.
  • A recurrent layer over a clip has no view of the whole match.

How it plugs into possession

As the trainable spotter when you own labels and need to ship. A permissive licence, a small model and frame-level outputs are the properties that matter when event evidence is one input to a larger possession system.

Its per-frame scores are also the right shape for a decoder. Do not threshold them into discrete events. Pass the score curves through as soft evidence of a touch.

Our read

The clean, permissive baseline for frame-accurate spotting. Bring your own football labels.

Sources

  1. spot: code for Spotting Temporally Precise, Fine-Grained Events in Video (ECCV 2022)github.com
  2. e2e-spot-models: released model weightsgithub.com