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
Original extreme occlusion paper:
- Robustness of Object Recognition under Extreme Occlusion in Humans and Computational Models
- https://arxiv.org/pdf/1905.04598.pdf
Building models that function similar to the brain:
- CORnet: Modeling the Neural Mechanisms of Core Object Recognition
- https://www.biorxiv.org/content/biorxiv/early/2018/09/04/408385.full.pdf
Importance of recurrence in occluded object recognition:
- Beyond core object recognition: Recurrent processes account for object recognition under occlusion.
- https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1007001