result doesn't match the paper

[iTomxy]

I prepare the data as needed (also the test set & retrieval set) and run the code almost as it is (only some logging & testing code added), and find it doesn't re-produce the result in the paper, getting only about 0.7128 of mAP@ALL.

• as evaluating codes are not provided, I use the one released at jiangqy/DCMH-CVPR2017.
• the way I prepare flickr can be found here. I use the label processed by myself to ensure a consistent class order in label vectors between those 3 splitting parts, but images are left untouched, i.e., they are the original raw images.

So how can the results in the paper be re-produced? are there any details I missed?

[jmojoo]

I have a similar issue. I ran experiments on NUSWIDE (prepared exactly as described in the paper). The MAP is almost the same (0.7646) but the WAP is different. The highest I got was 1.2507. I wrote my own evaluation code so maybe that's where the problem is, but I implemented it following the equations provided in the paper. If possible, could you provide some evaluation code?