A dataset for one-shot learning based contour of intereset extraction. Contact with qinfangbo2013@ia.ac.cn if you have any questions about the dataset.
This folder contains the ROCM dataset and the MF-ODS evaluation scripts.
- ROCM Dataset
The robotic object contour measurement (ROCM) datatset is built for evaluating the performance of ONE-SHOT learning based contour primitive of interest (CPI) extraction, as presented in the ICRA 2021 paper [1].
-
MF-ODS evaluation scripts: Since CPI map is a special case of edge/contour map, we follow the edge/contour/boundary detection works[2] and use the MF-ODS score to evaluate the model performance.
-
The dataset folder is arranged as follows:
|--- readme.txt
|--- test-pairs
|--- query-images (used for testing)
|--- object_*_*.jpg
|--- query-labels (used as CPI groundtruths)
|--- object_*_*_line/arc_*.png
|--- support-images (used for support branch, 1st-frame mode)
|--- object_*_1.jpg
|--- support-labels (CPI annotation)
|--- object_*_1_line/arc_*.png
|--- support-masks (object foreground annotation)
|--- object_*_1.png
|--- support-labels-extra (used for support branch, cross-device mode)
|--- object_*_1.jpg
|--- support-images-extra (CPI annotation)
|--- object_*_1_line/arc_*.png
|--- support-masks-extra (object foreground annotation)
|--- object_*_1.png
Example:
To test on the circular CPI on object 1 in the query image /query-images/object_1_5.jpg,
we use /query-labels/object_1_5_arc_0.png as the CPI GT,
then,
use /query-images/object_1_5.jpg as the query input,
use /support-images/object_1_1.jpg as the support image,
use /support-labels/object_1_1_arc_0.png as the support CPI annotation,
use /support-masks/object_1_1.png as the support object foreground annotation,
to predict the CPI image /test-rsts/object_1_5_arc_0.png using one-shot learning based CNN model.
To evaluate the accuracy of CPI extraction, we compare /query-labels/object_1_5_arc_0.png and /test-rsts/object_1_5_arc_0.png using the MF-ODS evaluation scripts.
Note that the folders with the suffix '-extra' and the folders without the suffix '-extra' only differ in how the support image is captured, which correspond to the '1st frame mode' and 'cross-device mode'. More details is found in [1].
- The MF-ODS evaluation scripts are provided in
|--- eval
|--- eval-extras
Configure the paths in demoBatchEval.m and run it with Matlab.
Note that this code is based on [2], and the misalignment tolerance parameter is set in correspondPixels.cc. If you need modify it, you need to re-compile correspondPixels.cc to correspondPixels.mexa64 using build.sh.
[1] Contour Primitive of Interest Extraction Network Based on One-Shot Learning for Object-Agnostic Vision Measurement, ICRA 2021, https://arxiv.org/abs/2010.03325