How Well are the Blind Baselines?

Published: Sep 1, 2020 by

Blind Baselines Perform Unexpectedly Well

We built three blind baselines which never use videos for training or inference. Our baselines put scores of deep models in a context. Surprisingly, our blind baselines are competitive and even outperform some deep models.

  • Prior-Only Blind : This baseline predicts temporal locations without using videos or query sentences.
  • Action-Aware Blind: This baseilne uses only one word in a query sentence to predict temporal locations of moments. For simplicity, we use the first verb in a query sentence.
  • Blind-TAN: This is a neural network-based model that uses the full query sentence. Blind-TAN is built upon 2D-TAN 1. Blind-TAN removes a module to extract visual features and replace the module with a learnable map to capture language biases. We trained Blind-TAN to predicts temporal locations solely with query sentences.

The score of TripNet is updated according to the latest published scores.

  1. Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2D temporal adjacent networks for moment localization with natural language. In The AAAI Conference on Artificial Intelligence, 2020. ↩

Share