pose_estimation

This code implements a simple framework for articulated pose estimation. In the current stage, the experiments are restricted to LSP dataset.

Generating clustering assignment

python cluster_relation.py --data ./joints_scaled_cropped.mat --human_exp_dir xxx --num_cluster 11

Training ResNet baseline on individual joints

python main.py --checkpoint_dir ./log_baseline --batch_size 32 --lr 1e-3 --joint_size 60

Training Side-by-side model

After training the baseline on individual joints, call

python main_pair.py --checkpoint_dir ./log_side_by_side --joint_size 60 --batch_size 32 --lr 1e-2 --pretrained ./log_baseline/model_best.pth --freeze_backbone 1 --hidden_size 100

Training Relational model

First training the (partial) model on individual joints:

python main.py --checkpoint_dir ./log_rel_individual --batch_size 32 --lr 1e-3 --joint_size 60 --use_rel 1

python main_pair.py --checkpoint_dir ./log_side_by_side --joint_size 60 --batch_size 32 --lr 1e-2 --pretrained ./log_rel_individual/model_best.pth --freeze_backbone 1 --hidden_size 100 --use_rel 1 --use_pos 1

Evaluating Models on individual joints

python main.py --mode eval --weights ./log_xxx/model_best.pth --batch_size 32 --lr 1e-3 --joint_size 60

Note that use_rel should be set accordingly.

Evaluating Models on target-cue pairs

python main_pair.py --weights ./log_xxx/model_best.pth --joint_size 60 --batch_size 32 --lr 1e-2 --pretrained ./log_rel_individual/model_best.pth --freeze_backbone 1 --hidden_size 100

Note that use_rel, use_pos should be set accordingly. There are also three arguments for ablation study, namely:

--ablation: for testing with 65 images selected for human experiments.
--rotate: for testing with different degrees of rotation on cue.
--cue_size: for testing with a cue size different from the target one.