setup.md

Setup annotations and data directories

Download template models and annotations

Download from here We provide rigged version of 3D template models for articulation. Every model has the following structured

cachedir
└───models
│   └───horse
│       │   mean_shape.mat ## contains mapping from UV value to faceindex on the template shape 
│       │   model.obj ## Base template shape
│       │   kp2vertex.txt ## Approximate 3D locations for keypoint vertices
│       │   mirror_transforms.txt ## Correspondence between transformation on reflection
│       │   hierarchy.xml ## Part hierarcy for articulation
│       │   parts.pkl ## Labelled vertices for every part
│       │   part_names.txt ## Part names
│   
└───bird
    │    ...

Download image annotations and splits

Download our pretrained model and cached annotations from here

cd acsm
tar -xf cachedir.tar

Download models for other imagenet catgories

cd acsm/cachedir/
wget https://www.dropbox.com/s/05lohn7x96o3fuf/models.zip?dl=0
unzip -q models.zip 

Training and Testing on CUBS dataset

• Train Birds with Keypoints. Generate training command using this

    python -m acsm.experiments.job_script --category=bird --kp=True --parts_file=acsm/part_files/bird.txt
  

• Train Birds without Keypoints

    python -m acsm.experiments.job_script --category=bird --kp=False --parts_file=acsm/part_files/bird.txt
  

• Evaluate KP Projection

  python -m acsm.benchmark.pascal.kp_project --name=acsm_bird_3parts --category=bird --parts_file=acsm/part_files/bird.txt --use_html --dl_out_pascal=True --dl_out_imnet=False --split=val --num_train_epoch=200 --num_hypo_cams=8 --env_name=acsm_bird_3parts_pck_val --multiple_cam=True  --visuals_freq=5 --visualize=True --n_data_workers=4 --scale_bias=1.5  --resnet_style_decoder=True --resnet_blocks=4 --el_euler_range=90 --cyc_euler_range=60
  

• Evaluate KP PCK Transfer

  python -m acsm.benchmark.pascal.kp_transfer --name=acsm_bird_3parts --category=bird  --parts_file=acsm/part_files/bird.txt --use_html --dl_out_pascal=True --dl_out_imnet=False --split=val --num_train_epoch=200 --num_hypo_cams=8 --env_name=acsm_bird_3parts_transfer_pck_val --multiple_cam=True --num_eval_iter=10000 --visuals_freq=1000  --visualize=True --n_data_workers=4  --scale_bias=1.5  --resnet_style_decoder=True  --resnet_blocks=4 --el_euler_range=90 --cyc_euler_range=60
  

Training and Testing on Pascal Horses dataset

• Train Horses with Keypoints. Generate training command using this

    python -m acsm.experiments.job_script --category=horse --kp=True --parts_file=acsm/part_files/horse.txt
  

• Train Horses without Keypoints

    python -m acsm.experiments.job_script --category=horse --kp=False --parts_file=acsm/part_files/horse.txt
  

• Evaluate KP PCK

  python -m acsm.benchmark.pascal.kp_project --name=acsm_horse_8parts --category=horse --parts_file=acsm/part_files/horse.txt --use_html --dl_out_pascal=True --dl_out_imnet=False --split=val --num_train_epoch=200 --num_hypo_cams=8 --env_name=acsm_horse_8parts_pck_val --multiple_cam=True  --visuals_freq=5 --visualize=True --n_data_workers=4 --scale_bias=0.75 --resnet_style_decoder=True --resnet_blocks=4 --el_euler_range=20 --cyc_euler_range=20
  

• Evaluate KP Projection

  python -m acsm.benchmark.pascal.kp_transfer --name=acsm_horse_8parts --category=horse  --parts_file=acsm/part_files/horse.txt --use_html --dl_out_pascal=True --dl_out_imnet=False --split=val --num_train_epoch=200 --num_hypo_cams=8 --env_name=acsm_horse_8parts_transfer_pck_val --multiple_cam=True --num_eval_iter=10000 --visuals_freq=1000  --visualize=True --n_data_workers=4  --scale_bias=0.75  --resnet_style_decoder=True  --resnet_blocks=4 --el_euler_range=20 --cyc_euler_range=20
  

Other configurations of models.

Model	Keypoint Supv	Num of Parts
acsm_bird_kp_3parts	Yes	3
acsm_bird_3parts	No	3
acsm_bird_kp_0parts	Yes	0
acsm_bird_0parts	No	0
acsm_horse_kp_8parts	Yes	8
acsm_horse_8parts	No	8
acsm_horse_kp_0parts	Yes	0
acsm_horse_0parts	No	0