/IBSR-ViewSequenceLearning

Primary LanguagePython

Readme

The implementation of Cross-Domain Image-Based 3D Shape Retrieval by View Sequence Learning.

Lee, Tang, et al. "Cross-Domain Image-Based 3D Shape Retrieval by View Sequence Learning." 2018 International Conference on 3D Vision (3DV). IEEE, 2018.

Usage

Files descriptions：

  1. globals.py: experiment configs.
  2. input.py: a small lib to load image / 3D shape files.
  3. run_eval_exp.sh: script for running experiments.

Prepare dataset

  1. Download dataset here.
  2. unzip it into the repository directory.
unzip IM2MN.zip

train

mkdir tmp

Prepare pretrained model:

./prepare_pretrained_alexnet.sh

train with pretrained model:

python train.py --caffemodel ./alexnet_imagenet.npy  --train_dir tmp --learning_rate 0.01

fine-tune with checkpoint file:

python train.py --weights tmp/model.ckpt-10000 --train_dir tmp --learning_rate 0.01

test

PHASE=val CKPTITER=15000 GPU=1 ./run_eval_exp.sh

This scripts does 4 steps for testing:

  1. extract features of images and 3D shapes. Save them into hkl files.
  2. retrieve. The results are saved into hkl files.
  3. evaluate. Print mAP.
  4. gen_retrievis.py. Generate result files for RetrieVis visualization.

visualize

git clone https://github.com/WeiTang114/RetrieVis
cd RetrieVis/
python simple.py <ibsr-triplet>/tmp/visresult/result.test.pool5.15000 --port <port>

Connect to http://<ip>:<port> to see the results。