Skip to content

Latest commit

 

History

History
218 lines (156 loc) · 10.3 KB

README.md

File metadata and controls

218 lines (156 loc) · 10.3 KB
Raw

ARCTIC baselines

Table of content

Overview

PyTorch Lightning: To avoid boilerplate code, we use pytorch lightning (PL) 2.0.0 to handle the main logic for training and evaluation. Feel free to consult the documentation, should you have any questions.

Comet logger: To keep track of experiments and visualize results, our code logs experiments using comet.ml. If you wish to use own logger service, you mostly modify the code in common/comet_utils.py. This code is only meant as a guideline; you are free to modify it to whatever extent you deem necessary.

To configure the comet logger, you need to first register an account and create a private project. An API code will be provided for you to log the experiment. Then you export the API code and the workspace ID:

export COMET_API_KEY="your_api_key_here"
export COMET_WORKSPACE="your_workspace_here"

It might be a good idea to add these commands to your ~/.bashrc file, so you don't have to load the environment every time you login to your machine. Add these lines to the end of ~/.bashrc.

We call the allocentric split and the egocentric split in our CVPR paper the p1 split and p2 split respectively.

Each experiment is tracked with a 9-character ID. When the training procedure starts, a random ID (e.g., 837e1e5b2) is assigned to the experiment and a folder (e.g., logs/837e1e5b2) to save information on this folder.

Getting started

Here we assume you have:

  1. Finished setting up the environment.
  2. Downloaded and unzipped the data following docs/data/README.md.
  3. Finished configuring your logger.

To use the data, you need to move them from unpack:

mv unpack data

You should have a file structure like this after running ls data/*:

data/arctic_data:
data  models

data/body_models:
mano  smplx

To sanity check your setup, lets run two dry-run training procedures. We call the allocentric split and the egocentric split in our CVPR paper the p1 split and p2 split respectively.

# train ArcticNet-SF on a very small ARCTIC dataset (allocentric split)
python scripts_method/train.py --setup p1 --method arctic_sf -f --mute --num_epoch 2
# train Field-SF on a very small ARCTIC dataset (egocentric split)
python scripts_method/train.py --setup p2 --method field_sf -f --mute --num_epoch 2

Now you enable your comet logger, by removing --mute to start logging to comet server. A url will be generated so that you can visualize the prediction of your model in training, and its corresponding groundtruth:

python scripts_method/train.py --setup p1 --method arctic_sf -f --num_epoch 2

Click on any example in Graphics of your comet experiment. If you can see that the hands and objects are in the left column, the groundtruth (mostly) is overlaid on the image, and the training finished 2 epochs, then your environment is in good shape.

As a first-timer, it is normal to have some issues. See our FAQ to see if there is a solution.

Training

⚠️ As per our CVPR paper, our evaluation protocol requires models to be trained only on the training set. You may not train on the validation set. You may use the validation set for hyper parameter search. The test set groundtruth is hidden for online evaluation. Here we provide instructions to train on the training set and evaluate on the validation set.

Here we detail some options in the argparse parsers we use. There are other options in the argparser. You can check src/parsers for more details.

  • -f: Run on a mini training and validation set (i.e., a dry-run).
  • --mute: Do not create a new experiment on the remote comet logger.
  • --setup: name of the split to use. A split is a partition of ARCTIC data into training, validation and test sets.
  • --trainsplit: Training split to use
  • --valsplit: Split to use to evaluate during training

The following code trains the single-frame allocentric baselines in our paper in the allocentric setting (i.e., the p1 split). For the complete training guide for all models, please refer to train.md

# training on ArcticNet-SF in the allocentric setting in our paper
python scripts_method/train.py --setup p1 --method arctic_sf --trainsplit train --valsplit minival
# training on InterField-SF in the allocentric setting in our paper
python scripts_method/train.py --setup p1 --method field_sf --trainsplit train --valsplit minival

Since you have the groundtruth in this split, you can view your metrics and losses under Panels of your comet experiment. You can also see the visualization of the prediction and groundtruth under Graphics in your comet experiment.

Evaluation

Here we show evaluation steps using our pre-trained models. Copy our pre-trained models to ./logs:

cp -r data/arctic_data/models/* logs/

Our evaluation process consists of two steps: 1) dumping predictions to disk (extract_predicts.py); 2) evaluating the prediction against the GT (evaluate_metrics.py).

Here we assume you are using the p1 or p2 splits and we show instructions for local evaluation.

⚠️ Instructions for p1 and p2 evaluation for the test set will be provided once the evaluation server is online.

Evaluate ArcticNet

Evaluate allocentric ArcticNet-SF on val set on p1 split:

# PLACEHOLDERS
python scripts_method/extract_predicts.py --setup p1 --method arctic_sf --load_ckpt logs/3558f1342/checkpoints/last.ckpt --run_on val --extraction_mode eval_pose
python scripts_method/evaluate_metrics.py --eval_p logs/3558f1342/eval --split val --setup p1 --task pose

Evaluate allocentric ArcticNet-LSTM on val set on p1 split:

# PLACEHOLDERS
python scripts_method/extract_predicts.py --setup p1 --method arctic_lstm --load_ckpt logs/423c6057b/checkpoints/last.ckpt --run_on val --extraction_mode eval_pose
python scripts_method/evaluate_metrics.py --eval_p logs/423c6057b/eval --split val --setup p1 --task pose

Evaluate egocentric ArcticNet-SF on val set on p2 split:

# PLACEHOLDERS
python scripts_method/extract_predicts.py --setup p2 --method arctic_sf --load_ckpt logs/28bf3642f/checkpoints/last.ckpt --run_on val --extraction_mode eval_pose
python scripts_method/evaluate_metrics.py --eval_p logs/28bf3642f/eval --split val --setup p2 --task pose

Evaluate egocentric ArcticNet-LSTM on val set on p2 split:

# PLACEHOLDERS
python scripts_method/extract_predicts.py --setup p2 --method arctic_lstm --load_ckpt logs/40ae50712/checkpoints/last.ckpt --run_on val --extraction_mode eval_pose
python scripts_method/evaluate_metrics.py --eval_p logs/40ae50712/eval --split val --setup p2 --task pose

Evaluate InterField

Evaluate allocentric InterField-SF on val set on p1 split:

# PLACEHOLDERS
python scripts_method/extract_predicts.py --setup p1 --method field_sf --load_ckpt logs/1f9ac0b15/checkpoints/last.ckpt --run_on val --extraction_mode eval_field
python scripts_method/evaluate_metrics.py --eval_p logs/1f9ac0b15/eval --split val --setup p1 --task field

Evaluate allocentric InterField-LSTM on val set on p1 split:

# PLACEHOLDERS
python scripts_method/extract_predicts.py --setup p1 --method field_lstm --load_ckpt logs/546c1e997/checkpoints/last.ckpt --run_on val --extraction_mode eval_field
python scripts_method/evaluate_metrics.py --eval_p logs/546c1e997/eval --split val --setup p1 --task field

Evaluate egocentric InterField-SF on val set on p2 split:

# PLACEHOLDERS
python scripts_method/extract_predicts.py --setup p2 --method field_sf --load_ckpt logs/58e200d16/checkpoints/last.ckpt --run_on val --extraction_mode eval_field
python scripts_method/evaluate_metrics.py --eval_p logs/58e200d16/eval --split val --setup p2 --task field

Evaluate egocentric InterField-LSTM on val set on p2 split:

# PLACEHOLDERS
python scripts_method/extract_predicts.py --setup p2 --method field_lstm --load_ckpt logs/701a72569/checkpoints/last.ckpt --run_on val --extraction_mode eval_field
python scripts_method/evaluate_metrics.py --eval_p logs/701a72569/eval --split val --setup p2 --task field

For details of extract_predicts.py, see here.

Visualization

We will use scripts_method/visualizer.py to visualize model performance and its corresponding GT. Unlike scripts_data/visualizer.py, here we shows the actual model input and output. Therefore, images are not full-resolution.

Options for visualizer.py:

  • --exp_folder: the path to the experiment; the prediction will be saved there after the extraction.
  • --seq_name: sequence to visualize. For example, s03_box_grab_01_1 denotes the sequence s03_box_grab_01 and camera id 1.
  • --mode: defines what to visualize; {gt_mesh, pred_mesh, gt_field_r, gt_field_l, pred_field_r, pred_field_l}
  • --headless: headless rendering. It renders the mesh2rgb overlay video. You can also render other modalities such as segmentation masks.

Examples to visualize pose estimation:

# PLACEHOLDERS
# dump predictions
python scripts_method/extract_predicts.py --setup p1 --method arctic_sf --load_ckpt logs/3558f1342/checkpoints/last.ckpt --run_on val --extraction_mode vis_pose

# visualize gt_mesh
python scripts_method/visualizer.py --exp_folder logs/3558f1342 --seq_name s03_box_grab_01_1 --mode gt_mesh

# visualize pred_mesh
python scripts_method/visualizer.py --exp_folder logs/3558f1342 --seq_name s03_box_grab_01_1 --mode pred_mesh

Examples to visualize interaction field estimation:

# PLACEHOLDERS
# dump predictions
python scripts_method/extract_predicts.py --setup p1 --method field_sf --load_ckpt logs/1f9ac0b15/checkpoints/last.ckpt --run_on val --extraction_mode vis_field

# visualize gt field for right hand
python scripts_method/visualizer.py --exp_folder logs/1f9ac0b15 --seq_name s03_box_grab_01_1 --mode gt_field_r

# visualize predicted field for left hand
python scripts_method/visualizer.py --exp_folder logs/1f9ac0b15 --seq_name s03_box_grab_01_1 --mode pred_field_l

For details of extract_predicts.py, see here.