search

AI-NERC-NUPT / PLR-OSNet

Source codes for the paper "Learning Diverse Features with Part-Level Resolution for Person Re-Identification"
44 stars 18 forks source link

Some question about map 80% on cuhk03-labeled dataset #3

Open HouBiaoLiu opened 4 years ago

HouBiaoLiu commented 4 years ago

Experiment 1: I download the cuhk03-np datasets from https://github.com/zhunzhong07/person-re-ranking/tree/master/CUHK03-NP. Then I trained cuhk03_labeled, I got the best map 57%.

Experiment 2:

First：download cuhk-03.mat and the new training/testing protocol split cuhk03_new_protocol_config_detected.mat cuhk03_new_protocol_config_labeled.mat, place all in the directory '/data/person-reid/cuhk03/'

Second: run python command from torchreid.data.datasets.image import cuhk03 CUHK03('/data/person-reid/cuhk03/') Now I get splits_new_detected.json and splits_new_labeled.json in the same directory '/data/person-reid/cuhk03/'

Third: genarate these folders bounding_box_train as followes: from torchreid.utils import mkdir_if_missing, read_json, write_json import shutil

split_new_det_json_path = '/data/person-reid/cuhk03/splits_new_detected.json'

            split_new_lab_json_path = '/data/person-reid/cuhk03/splits_new_labeled.json'
            splits = read_json(split_new_lab_json_path)
            num = 0
            for item in splits[0]['train']:
                    pid = item[1]
                    cam = item[2]
                    filename = '%04d' % pid +'_'+'c'+'%d' % (cam+1)+'_'+'%04d' % num + '.png'
                    num = num +1
                    print (item, filename)
                    shutil.copyfile(item[0], os.path.join('/data/person-reid/cuhk03/CUHK03_labeled/bounding_box_train', filename))

Now I am training on CUHK03_labeled, looking forward to the result

The question is how to generate the bounding_box_train query bounding_box_test from the raw data cuhk-03.mat. Do all methods evaluate on the new training/testing protocol split. Then we can train as your suggest and reproduce your siginificant result

HouBiaoLiu commented 4 years ago

on the new training/testing protocol split, finally I train done and get the map 54.6%. Can you tell me how to generate the bounding_box_train query bounding_box_test from the raw data cuhk-03.mat, Is my way wrong?