search

csuhan / ReDet

Official code of the paper "ReDet: A Rotation-Equivariant Detector for Aerial Object Detection" (CVPR 2021)
https://redet.csuhan.com
Apache License 2.0
387 stars 79 forks source link

**important**whats happening in prepare_dota1 whats the error for testing #83

Open mmoghadam11 opened 3 years ago

mmoghadam11 commented 3 years ago

hi @csuhan thanx for your implementation but i have one important error everywhere here and in AerialDetection both one:

  1. i use prepare_dota1 and got dota1_1024 folder and .json files
  2. download .pth files and make data and work_dir folders & put the addresses in configs and got 3 folder for each config & use task1_results_nms.zip for evaluating server

i do everything that said in GETTING_STARTED.md here and AerialDetection but every time i got mAP like this : AerialDetection : mAP: 0.3745174540144275 ap of each class: plane:0.5300541025955471, baseball-diamond:0.27181818181818185, bridge:0.43130126717635453, ground-track-field:0.14171122994652408, small-vehicle:0.24575509050331162, large-vehicle:0.25701370866991136, ship:0.7015457849808598, tennis-court:0.5453346126169345, basketball-court:0.3152847152847153, storage-tank:0.09090909090909091, soccer-ball-field:0.18932806324110674, roundabout:0.1655011655011655, harbor:0.6844767188033766, swimming-pool:0.6143607552420431, helicopter:0.4333673229272884

ReDet_re50_refpn_1x_dota1.py: mAP: 0.3796692927779224 ap of each class: plane:0.5393134760132164, baseball-diamond:0.2727272727272727, bridge:0.4428214528237172, ground-track-field:0.14898989898989898, small-vehicle:0.2509144865918008, large-vehicle:0.25942160410103354, ship:0.7073533101735577, tennis-court:0.5454545454545455, basketball-court:0.32725572697522204, storage-tank:0.09090909090909091, soccer-ball-field:0.21752528974341798, roundabout:0.14793388429752066, harbor:0.6952776159604523, swimming-pool:0.6333621700100804, helicopter:0.41577956689800866

i use google colab tesla t4 what happening why APs are low????

But: today i use dota_evaluation_task1.py and use dota valset for evaluating and got this: using ReDet_re50_refpn_1x_dota1.py npos num: 72 ap: 0.8435975774428759 map: 0.8514600172670281 classaps: [90.74063962 88.35952404 70.27778167 83.69586216 71.37892832 88.03846396 88.83972303 90.90909091 89.87234694 90.00746689 90.00924415 82.27596327 88.32895278 80.09628041 84.35975774]

so it seems the network works good so what is the error in testing??? maybe prepare_dota1 doesnt work good???

csuhan commented 3 years ago

Can you provide the log (console output)?

mmoghadam11 commented 3 years ago

Can you provide the log (console output)?

log for testing?? i dont understand what do you mean?? it just test 4750 splitted picture and make a .pkl file

mmoghadam11 commented 3 years ago

after prepare_dota1 i have 4750 splitted img for testing and a .json file that named DOTA_test1024.json for splitted img : https://drive.google.com/file/d/1-AgdUjDYlHC1MueH1spX3ixbTuwUmI-k/view?usp=sharing

csuhan commented 3 years ago

There should be 937/10833 testing images before/after spliting. Please check your data.

mmoghadam11 commented 3 years ago

There should be 937/10833 testing images before/after spliting. Please check your data.

in dota v1??? testing folder has 937 images??? what about train an val number images???

csuhan commented 3 years ago

There are 2806 images. See our paper: DOTA: A Large-scale Dataset for Object Detection in Aerial Images