search

JXingZhao / ContrastPrior

The Code of Contrast Prior and Fluid Pyramid Integration for RGBD Salient Object Detection(CVPR2019)
51 stars 8 forks source link

the data result #1

Open Mrlong12 opened 5 years ago

Mrlong12 commented 5 years ago

why your stereo dataset result have 1000 images? other papers only have 797 images like that, image but yours is like that image

JXingZhao commented 5 years ago

This dataset is relatively early and not widely used. We ask the author of the dataset directly for the images, so it is correct. As for the 797 images used in some articles, because some researchers have lost some images of the dataset, the follow-up work is mostly done on the 797 dataset.

Mrlong12 commented 5 years ago

And in the NJU2000 dataset ,the result you gave only has 485 images like that: image

but other paper has 503 images like that image

yangcaoai commented 5 years ago

And in the NJU2000 dataset ,the result you gave only has 485 images like that: image

but other paper has 503 images like that image

Hi, you can check the 'readme.txt' in NJU2k： `NJUDS2000--A dataset designed for evaluation of depth-aware salient object detection methods. This dataset includes more than 2000 stereo images, as well as depth maps and manually labeled ground truth.

--LR-- This folder includes the original color images. The left and right images are named with suffixes of "_left" and "_right" respectively. All the images have been rectified to meet the epipolar constraint.

--GT-- This folder contains the ground truth salient object masks of the left images. The masks are recorded in binary images, where "1" indicates object and "0" stands for background.

--depth-- This folder consists of the depth maps of the left images, which are generated by [1]. Each image is rescaled to [0,255], the larger intensity, the nearer pixel to observer. Please note that some of the depth maps could be inaccurate and thus may puzzle your method. One can also make use of the other stereo methods to get more accurate depth results.

[1] D. Sun, S. Roth, M. J. Black, Secrets of optical flow estimation and their principles, in: IEEE Conference on Computer Vision and Pattern Recognition, IEEE, 2010, pp. 2432šC2439.

Release Notes:

Version 1. 2014.10.23 2000 images. 400 of them are chosen as a trial dataset: NJUDS400.

Version 2. 2015.6.18 1985 images. 97 images with bad depth maps are removed. Then 100 new images are carefully chosen to add to the dataset. At last, the left and right color images of 18 personal photos are removed for the protection of individual privacy. However, the depth maps and ground truth masks are retained.`

As shown above，the dataset of version2 has 1985 images. For fair comparison, our train/val set keep the same number of images with PCF. Then, the test set has 485 images.