Baseline for SLRTP Track 3: Spotting on BSL Corpus

This repository is intended to produce a baseline method for spotting on BSL Corpus. Using the provided logits in the data for this challenge, the model computes the top-5 predictions at each frame. Given a query gloss, we consider a sign gloss for this query to be localised if amongst the top-5 predictions.

The structure of this repository also contains a dataloader to load model features and labels, and a simple MLP model that can be trained on the features.

To evaluate baseline model using the provided logits on test set:

bash commands/test.sh

This create a file called res/submission_dev.zip, which may be uploaded to the CodaLab server.

To train a simple MLP model

bash commands/train.sh