vpulab/urbam-reid

UrbAM-ReID Dataset

UrbAM-ReID is a long-term geo-positioned urban ReID dataset. It is composed by four subdatasets recording the same trajectory at the UAM Campus, each one recorded in different seasons and including an inverse direction recording.  This work specifically addresses urban objects re-identification, currently, waste containers, rubbish bins, and crosswalks. The dataset provides different attributes of the annotated objects, like their classes, their foreground or background status and the geo-position. Several evaluation configurations can be defined to simulate realistic scenarios that may arise in actual situations within the management of urban elements, considering the utilization of just visual data, or incorporating additional attributes, providing different complexity levels. Finally, the dataset is used for defining a benchmark where two open source state-of-the-art systems are evaluated.

link of UrbAM-ReID dataset

Link to the dataset and code

Supplementary material

It is available in Supplemetary_material.pdf

Code

This porposal utilizes two open-source state-of-the-art works:

1. Part-Aware Transformer
2. ReID Strong Baseline

Modified Codes

The folder includes modified codes to integrate the UrbAM-ReID dataset into each system.

Evaluation

To evaluate the system, follow the instructions in each GitHub repository. After training, use the update.py (or update_GPS.py for GPS post-processing) script to infer the data. Evaluate the obtained track.txt results using Evaluate_UrbAM-ReID.py.

Example of BOT update.py

python update.py --config_file=./configs/SOA/softmax_triplet_UAM_test_cross1.yml --track=./checkpoints/crosswalk/crosswalk13new/track2_model1UAM_tradicional.txt MODEL.DEVICE_ID "('0')" DATASETS.NAMES "('UAM_test')" TEST.NECK_FEAT "('after')" TEST.FEAT_NORM "('yes')" MODEL.PRETRAIN_CHOICE "('self')" TEST.RE_RANKING "('yes')" TEST.WEIGHT "('./checkpoints/crosswalk/crosswalk13new/resnet50_model_100.pth')"

Example of BOT update_GPS.py

python update_GPS.py --config_file=./configs/SOA/softmax_triplet_UAM_test_cross1.yml --track=./checkpoints/crosswalk/crosswalk13new/track2_model1UAM_tradicionalGPS.txt --xml_dir_gallery=test_label_all.xml --xml_dir_query=query_label_all.xml MODEL.DEVICE_ID "('0')" DATASETS.NAMES "('UAM_test')" TEST.NECK_FEAT "('after')" TEST.FEAT_NORM "('yes')" MODEL.PRETRAIN_CHOICE "('self')" TEST.RE_RANKING "('yes')" TEST.WEIGHT "('./checkpoints/crosswalk/crosswalk13new/resnet50_model_100.pth')"

Example of PAT update.py

python update.py --config_file "config/UAM_cross1_test.yml" --track ./logs/UAM/UAM_cross1/track2_model1UAM_tradicional

Example of PAT update_GPS.py

python  update_GPS.py --config_file "config/SoA/UAM_cross1_test.yml" --track ./logs/UAM/UAM_cross1/track2_model1UAM_tradicionalGPS --xml_dir_gallery test_label_all.xml --xml_dir_query query_label_all.xml

GPS Annotations

The ./GPSannotations directory is divided into four subfolders, each related to a sequence (March, May, July, and July inverse). Each subfolder contains a CSV file with GPS annotations.

• The first row is a header indicating the columns: frame, lat, long, azi.
• The subsequent rows of the file provide information on GPS coordinates per frame.
• GPS annotation is included in this proposal as a post-processing step, but the possibilities of use are extensive.

ReID Sequences

The ./ReIDSequences folders contain annotated files obtained from CVAT annotations.xml with all the attributes, and the images_output with the format XXXXXX_Y_Z.jpg, where XXXXX is the original frame name, Z is the annotated ID, and Y is the number of this identity in this location. In the case of containers, the same identity is considered per the same container class in the same location, and it enumerates the number of containers of the same class in this location.

Splits

The ./splits directory collects five different splits per urban object.

• For each split, it has the baseline data (./image_train, ./image_test, image_query).
• In ./FG, there are just the foreground annotated objects for the baseline.
• In ./inv, it contains the baseline data adding the inverse as a query (the baseline query now is in image_test).
• In ./inv/FG, there are just the foreground annotated objects for the data including the inverse.

Each data scenario includes the following files with annotations:

• train_label.xml, test_label.xml, query_label.xml: Includes cameraID (each ID=1 is May sequence, ID=2 is March, ID=3 is July, and ID=4 is July inverse), image name related to this split, and object ID.
• train_label_all.xml, test_label_all.xml, query_label_all.xml: Includes the same cameraID; in this case, the imageName is the frame of the original sequence (the same as "frame" in GPS .csv, and in XXXXXX in image output, and the same as in ReIDSequences/../annotations.xml attribute image name). The rename is the image name in this folder (in the case of the FG folder, it also appears as rename2 due to it is related to the baseline folder).

Acknowledgment

This work has been supported by the Ministerio de Ciencia, Innovación y Universidades of the Spanish Government under project SEGA-CV (TED2021-131643A-I00)