about the cuhk03 evaluation

[zgplvyou]

thanks for your evaluation code , but i am a little confused about one line of it. could you tell me what 'ff = ff.ff1' and 'ff =ff.ff2' mean. i don't find where the 'ff , ff1' is.

[zgplvyou]

and why '[,index]= sort(score,'ascend')' in market1501 evaluation and '[,index]= sort(score,'descend')' in cuhk03 evaluation?

[layumi]

Hi @zgplvyou
ff1 is query feature
ff2 is gallery feature
We evaluate the result by the matching score. So we use sort

[zgplvyou]

i have used your market1501 evaluation code , so i know the 'sort' mean. and why sort by ‘ascend' in market1501 but 'descend' in cuhk03

[layumi]

In market1501, I followed the original code using (x-y)^2
In cuhk, I rewrite the code using x*y (x,y have been normalized)
since x^2 =1, y^2=1

[zgplvyou]

thanks for Zheng ge's answers!

[zgplvyou]

Sorry for one more question  about 'single-shot' and 'multi-shot'.
i know the 'single-query' and 'multi-query' due to your explanation. and i found the definition of 'single-shot' and 'multi-shot' that In general, when only an image pair is matched, the method is considered as single-shot recognition method. If matching is conducted between two sets of images, e.g. frames obtained from two separate trajectories. 

the test images of CUHK03 is obtained from 6 cameras(3 camera pairs[camera_1,camera_1'; camera_2,camera_2'; camera_3,camera_3']). the amount of test images is about 100 * 10.
so how to choose the query images and gallery images for single-shot and multi-shot.
is the prob=[all test images in camera_1, camera_2, camera_3], gallery = [random choose one image of each identity in camera_1', camera_2', camera_3'] in single-shot right? but in this case, the matching is obtained from different cameras. i am confused of the choice of probe and gallery on single-shot and multi-shot.

[layumi]

The original setting is single shot. Every time we select 1 query and 100 images from 100 person (another camera . For every person, we randomly selects one image) as gallery.
And run 20 times to report the mean precision.

Multi-shot is similar to market1501. We select 1 query and all images (another camera) as gallery.
Run one time.

[zgplvyou]

Thank you again! i have got the rank1, rank5 and map.

[daichenchao]

thanks for your evaluation code , but i am a little confused about one line of it. could you tell me what 'load('cuhk03_feature_res_split6.mat')' mean. i don't find where the 'cuhk03_feature_res_split6.mat' is.

[layumi]

Hi @daichenchao
You need to generate it. I wrote a prepare script.
You may check here.https://github.com/layumi/2016_person_re-ID/blob/master/cuhk03-prepare-eval/prepare_imdb.m#L9

[daichenchao]

Thanks for your reply! I have performed /cuhk03-prepare-eval/prepare_imdb.m I can not find cuhk03_feature_res_split6.mat rather than cuhk_data_256_split6.mat. cuhk03_feature_res_split6.mat is located in 2016_person_re-ID-master\cuhk03-prepare-eval\eval_cuhk03_10times.m. it is located in first line.I think this file should be stored in the characteristics of resnet-50 extraction

[daichenchao]

Sorry to bother you. I have found the cuhk03_feature_res_split6.mat generated file. Very embarrassing, I did not double-check.Thank you again! i have got the rank1, rank5 and map.

[rmxhhh]

sorry to bother you,can you tell me where is the' cuhk03_feature_res_split6.mat',I can not find it.Thank you very much. @daichenchao

[daichenchao]

a function file in document is used to generate cuhk03_feature_res_split6.mat.you can find it carefullly

[rmxhhh]

I found it ,Thank you! @daichenchao