Skip to content

Pytorch implementation of an Hourglass network for corner detection of a wrinkled towel.

Notifications You must be signed in to change notification settings

AlbertMitjans/pytorch-corner-detection

Repository files navigation

Hourglass for corner detection of deformable fabric

This repo contains the code structure used for the detection of the corners of a wrinkled towel.

Installation

Create conda environment

$ conda create -n ENVIRONMENT_NAME python=3
$ conda activate ENVIRONMENT_NAME

Clone and install requirements

$ git clone [email protected]:AlbertMitjans/pytorch-corner-detection.git
$ cd pytorch-corner-detection/
$ conda install --file requirements.txt

Run test

Evaluates the model on the dataset.

$ python3 main.py --train False --ckpt checkpoints/best_ckpt/model.pth

As default, for every case, the depth, the output, and the rgb are saved in the output/ folder.

Testing log

* Recall(%): 54.830     * Precision(%):  (97.727, 79.804, 41.202, 14.815)    

The precision is computed for the (1, 2, 3, 4) corners detected with highest confidence (the gaussians with a highest value on its center).

Run train

Trains the network from scratch or from a given ckpt.

$ python3 main.py

Training log

Epoch: [5][300/312]	Loss.avg: 0.3615	Recall(%): 21.622	Precision num. corners (%): (22.591, 18.563, 15.833, 16.809)

Tensorboard

Track training progress in Tensorboard:

  • Initialize training
  • Run the command below inside the pytorch-corner-detection directory.
  • Go to http://localhost:6006/
$ tensorboard --logdir='logs' --port=6006

Real-time display and image capture

  1. Watch the output of the network in real-time by connecting a depth camera.

  2. Capture images to increase the size of the dataset.

In order to work with the encoded images received by the camera, we need to convert them to OpenCV images using the cv_bridge package. In python 3, this package is not compatible with rospy, therefore I created another repository which implements this in python 2.

Arguments

--train (default:True) : if True/False, training/testing is implemented.
--val_data (default:True) : if True/False, all/validation data will be evaluated.
--save_imgs (default:True) : if True output images will be saved in the \Output folder.
--batch_size (default:1)
--depth (default:True) : if True/False, depth/RGB images will be used.
--ckpt(default:None)
--num_epochs (default:200)

About

Pytorch implementation of an Hourglass network for corner detection of a wrinkled towel.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages