train.md

Pre-reqs

Install external loss lib

git clone https://github.com/shubhtuls/PerceptualSimilarity.git

CUB Data

1. Download CUB-200-2011 images somewhere:

cd misc &&  wget http://www.vision.caltech.edu/visipedia-data/CUB-200-2011/CUB_200_2011.tgz && tar -xf CUB_200_2011.tgz

2. Download CUB annotation mat files and pre-computed SfM outputs. This should be saved in misc/cachedir directory:

wget https://sfo2.digitaloceanspaces.com/yun/misc/cachedir.tar.gz && tar -vzxf cachedir.tar.gz

Computing SfM

You may skip this

We provide the computed SfM. If you want to compute them yourself, run via matlab:

cd preprocess/cub
main

Model training

you need to start visdom.server before training

nohup python -m visdom.server & # remove nohup and `&` to see error

Change the name to whatever you want to call. Also see main.py to adjust hyper-parameters (for eg. increase tex_loss_wt and text_dt_loss_wt if you want better texture, increase texture resolution with tex_size). See nnutils/mesh_net.py and nnutils/train_utils.py for more model/training options.

python  main.py --name=bird_net --display_port 8097

More settings:

# Stronger texture & higher resolution texture.
python main.py --name=bird_net_better_texture --tex_size=6 --tex_loss_wt 1. --tex_dt_loss_wt 1. --display_port 8088

# Stronger texture & higher resolution texture + higher res mesh. 
python main.py --name=bird_net_hd --tex_size=6 --tex_loss_wt 1. --tex_dt_loss_wt 1. --subdivide 4 --display_port 8089

Evaluation

We provide evaluation code to compute the IOU curves in the paper. Command below runs the model with different ablation settings. Run it from one directory above the cmr directory.

python misc/benchmark/run_evals.py --split val  --name bird_net --num_train_epoch 500

Then, run

python misc/benchmark/plot_curvess.py --split val  --name bird_net --num_train_epoch 500

in order to see the IOU curve.