Download PASCAL3D+_release1.1.zip from the official website and extract it to datasets/p3d/. Download the P3D cache archive from the CMR repo (source) and extract it to datasets/p3d/. Finally, download poses_perspective.zip (source: textured-3d-gan repo) and extract it to datasets/. Your directory tree should look like this:
datasets/p3d/
datasets/p3d/PASCAL3D+_release1.1/
datasets/p3d/p3d_sfm_image/
datasets/p3d/p3d_car/detections.npy
datasets/p3d/p3d_car/poses_estimated_singletpl_perspective.bin
Download CUB_200_2011.tgz (source) and extract it to datasets/cub/. Then, download segmentations.tgz (source) and extract it to datasets/cub/CUB_200_2011/.
You also need to download the poses (source: CMR repo) and extract them so that your directory tree finally looks like this:
datasets/cub/
datasets/cub/data
datasets/cub/sfm
datasets/cub/CUB_200_2011/
datasets/cub/CUB_200_2011/segmentations/
To set up images from ImageNet, you need to download the synsets in the table below from the full ImageNet22k dataset (depending on the category you want to evaluate). The full dataset can be found on Academic Torrents.
| Category | Synsets |
|---|---|
| motorcycle | n03790512, n03791053, n04466871 |
| car | n02814533, n02958343, n03498781, n03770085, n03770679, n03930630, n04037443, n04166281, n04285965 |
| airplane | n02690373, n02691156, n03335030, n04012084 |
| elephant | n02504013, n02504458 |
| zebra | n02391049, n02391234, n02391373, n02391508 |
You should then copy the individual synset directories to datasets/imagenet/images/. You additionally need to download poses_perspective.zip (source: textured-3d-gan repo) and extract it to datasets/.
Your directory tree should eventually look like this:
datasets/imagenet/
datasets/imagenet/images/n*
datasets/imagenet/imagenet_*/detections.npy
datasets/imagenet/imagenet_*/poses_estimated_multitpl_perspective.bin
Download srn_chairs.zip and srn_cars.zip (source: PixelNeRF repo) and extract them to datasets/shapenet/.
Download carla.zip and carla_poses.zip (source: GRAF repo) and extract them to datasets/carla/.
If want to train the encoder from scratch, you first need to download the pretrained backbone weights mit_b5.pth (pre-trained on ImageNet) from the SegFormer model archive (source) and extract it to coords_checkpoints/.