README.md

Exploring Deep Closest Point: Learning Representations for Point Cloud Registration

Sample Run

# Example of how to run Training and Testing for color input on our Mixamo dataset

# Training 
python main.py --device cuda:0 --exp-name=color_input --model=dcp --emb-nn=dgcnn --pointer=transformer --head=svd --test-batch-size 5 --dataset mixamo --use-color True --arap True --model-path checkpoints/model_color_input.249.t7

# Test
python main.py --eval --device cuda:0 --exp-name=color_input --model=dcp --emb-nn=dgcnn --pointer=transformer --head=svd --test-batch-size 5 --dataset mixamo --use-color True --arap True --model-path checkpoints/model_color_diffsample.250.t7

Important parameters

--device cpu # XPU Device to run the model on (eg. cpu, cuda:0, cuda:1)
--use-color True # Flag for using the color as input
--different-sampling True # Flag for using different sampling in source and target
--arap True # Flag for ARAP regularizer
--num-points 1024 # "Num of points to use" 
--dataset mixamo # dataset to use one of the following [modelnet40, mixamo,tumrgbd]
--model-path checkpoints/model_color_input.249.t7 # Path for checkpoint / pretrained model. For training this parameter is used for continuing from a checkpoint. For test it is the path of the pre-trained model
--matching-method # The point matching method, one of the following ['softmax, 'sink_horn']
--num_sk_iter # Number of inner iterations used in sinkhorn normalization if sinkhorn is enabled

Report

For full report in PDF format click here

Overview

Note: Images are clickable for higher resolution

Page 1	Page 2

Page 3	Page 4	Page 5