search

Ben93kie / SeaDronesSee

Vision Benchmark for Maritime Search and Rescue
72 stars 25 forks source link

Data set inconsistency #11

Closed SuperXxts closed 2 years ago

SuperXxts commented 2 years ago

The official website（https://seadronessee.cs.uni-tuebingen.de/） mentioned that there were 5630 pictures in the training set of target detection tasks, but in the public data set（https://cloud.cs.uni-tuebingen.de/index.php/s/pNqRDqtLjXbEqDm）, I only found 2975 pictures . There are five categories mentioned in the paper, namely: swimmer, float, swimmer on boat, float on boat, boat, but there are six categories in the data set? So how many categories should there be?

Ben93kie commented 2 years ago

The webpage had a typo, we fixed that. Thanks!

In the original SeaDronesSee dataset we feature the following classes:

swimmer, swimmer on boat, floater, floater on boat, life jacket, boat.

In the synthetic counterpart (DGTAV-SeaDronesSee) we only feature the following classes:

swimmer, swimmer on boat, floater, floater on boat, boat.

Maybe you mean the synthetic dataset?

Ben93kie commented 2 years ago

Closed because of inactivity. Reopen if some issue persists.

FraCamp commented 1 year ago

Hi @Ben93kie (good job btw), in the SeaDronesSee.yaml you mention 6 different classes, which are different from the SDS paper: yaml -> names: [ 'ignored', 'swimmer', 'boat', 'jetski', 'life_saving_appliances', 'buoy'] while paper : ['swimmer', 'swimmer on boat', 'floater', 'floater on boat', 'life jacket', 'boat'] Can you please explain better these difference, and what version is used in the pretrained_model_on_SeaDronesSee_ObjectDetection_V2.pt

C-der-Baum commented 1 year ago

Hi, thanks @FraCamp! There are two versions of SeaDronesSee:

  1. The original SeaDronesSee Object Detection Dataset
  2. The updated version SeaDronesSee Object Detection v2 introduced in the MaCVi workshop (read here https://openaccess.thecvf.com/content/WACV2023W/MaCVi/html/Kiefer_1st_Workshop_on_Maritime_Computer_Vision_MaCVi_2023_Challenge_Results_WACVW_2023_paper.html)

The latter version has different classes, namely [ 'ignored', 'swimmer', 'boat', 'jetski', 'life_saving_appliances', 'buoy']. The object detector in this repository was trained on this dataset.

Does that answer your question?