/recurrent-extreme-occlusion

Investigating the role of recurrent connections in object recognition under extreme occlusion.

Primary LanguageJupyter Notebook

6.804 Computational Cognitive Science Final Project

Evaluation

Since CORnet-* output 1000 classes corresponding to the 1000 ImageNet classes, whereas the Vehicle Occlusion dataset (source: ) is based on PASCAL3D+, we need to map the 1000 classes to the corresponding six vehicle in the PASCAL3D+ dataset.

Using the 1000 class indices from ImageNet from https://gist.github.com/yrevar/942d3a0ac09ec9e5eb3a, we set the following mapping from PASCAL3D+ to ImageNet.

TODO: fill this in

Full Results Table

Google Spreadsheet

References

Original extreme occlusion paper:

Building models that function similar to the brain:

Importance of recurrence in occluded object recognition: