README

Here is a project of RGB-D Saliency Object Detection, we will share the most common datasets, tools of calculating evaluation metrics, some skills of pre-trained and our new research advances...

---

目录

• Datasets
• Evaluation_tools
• Pre-trained_tools

---

Datasets

There are eight mainstream RGB-D saliency datasets:

Dataset	DES/RGBD135	DUT-RGBD	LFSD	NJUD	NLPR	SIP	SSD	STEREO
Size	135	1200	100	2000	1000	929	80	1000/797

• Tips: The authors of NJUD dataset published 2000 samples initially, but they updated them to 1985 samples subsequently.
• Tips: The authors of STEREO dataset published 1000 samples initially, but they updated them to 797 samples subsequently.

There are two generally accepted settings of training datasets:

• Training dataset 1, including NJUD(1485) and NLPR(700), fetch code is 2185.
• Training dataset 2, including NJUD(1485) and NLPR(700), DUT-RGBD(800), fetch code is 2985.

The remaining samples in the above mentioned datasets are used for test:

• All test datasets, fetch code is test.
• We use 300 samples of NLPR, 400 samples of DUT-RGBD for test.
• We provide two versions of the STEREO dataset(797/1000) for test.

Tips: All data is resized to 256×256 pixels.

---

Evaluation_tools

• Put the test datasets and sal maps of other papers into the './evaluation_tool/' folder.
• Modify the 'salmap-root' and 'test_data', then run /evaluation_tool/calculate_metrics.py.
• Tips: The name of test dataset and sal maps shoule be the same.

---

Pre-trained_tools

• Tips: some datasets satisfy the rule that closer objects present lower depth values (are "black"), while further objects have higher depth values (are "white"). Although these depth maps are normalized into the range (0~1), such a rule is enforced to meet physical common sense. We provide the tools to inverse the black and white pixels of depth images.
• Run depth_inverse.py, you can inverse the black and white pixels of depth images.
• Run sal_to_edge.py to generate edge maps from the ground truth of mask.

---