Joint learning of object and action detectors

By Vicky Kalogeiton, Philippe Weinzaepfel, Vittorio Ferrari, Cordelia Schmid

Introduction

In our ICCV17 paper we jointly detect objects performing various avtions in videos. Here, we provide the evaluation code used in our experiments.

Citing Joint learning of object and action detectors

If you find our evaluation code useful in your research, please cite:

@inproceedings{kalogeiton17biccv,
  TITLE = {Joint learning of object and action detectors},
  AUTHOR = {Kalogeiton, Vicky and Weinzaepfel, Philippe and Ferrari, Vittorio and Schmid, Cordelia},
  YEAR = {2017},
  BOOKTITLE = {ICCV},
}

Evalutation of all architectures

You can download our detection results (multitask, hierarchical and cartesian):

curl http://pascal.inrialpes.fr/data2/joint-objects-actions/JointLearningDetections.tar.gz | tar xz 

To run the mAP evaluation function for the multitask, hierarchical and cartesian cases (Table 4 in our paper), run:

vic_A2D_eval(learning_case) # learning_case: 1, 2 or 3 for multitask, hierarchical and cartesian, respectively

Zero shot learning

To run the mAP evaluation function for the zero shot learning (Table 5 in our paper), run:

vic_A2D_eval_zeroshot 

Object-action semantic segmentation

You can download our semantic segmentation images:

curl http://pascal.inrialpes.fr/data2/joint-objects-actions/JointLearningSegmentationResults.tar.gz | tar xz 

To run the semantic segmentation evaluation function (Table 6 in our paper), run:

vic_eval_segmentation