search

Thinklab-SJTU / R3Det_Tensorflow

Code for AAAI 2021 paper: R3Det: Refined Single-Stage Detector with Feature Refinement for Rotating Object
Apache License 2.0
540 stars 122 forks source link

got mAP of 0.52 in dotav1? #125

Closed nishit01 closed 3 years ago

nishit01 commented 3 years ago

Hi, Thanks for open sourcing your code. I followed the steps for the setup

However, I am getting mAP of 0.52 for the training image in DOTAv1 dataset

Following are the steps I followed for testing the trained images and calculating mAP

  1. Downloaded the model - R3Det + all the pre-trained weights https://github.com/Thinklab-SJTU/R3Det_Tensorflow#download-model and placed under data/pretrained_weights
  2. Performed Testing on the train images by following the command in this section https://github.com/Thinklab-SJTU/R3Det_Tensorflow#test and placed img_dir pointing to the train images
  3. Results obtained under dotares (starting with Tas1) are then evaluated with dota_evaluation_task1.py from DOTA_devkit

Kindly let me know if I am making some mistakes and how to rectify it. Thanks

yangxue0827 commented 3 years ago

You need train the model if you noly use pretrain weight or you should download the trained model before directly test the image

nishit01 commented 3 years ago

Hi @yangxue0827 thanks for responding ... I have downloaded the trained model and placed under .../R3Det_Tensorflow/output/trained_weights/RetinaNet_DOTA_R3Det_4x_20200819 and this link is the one I used for trained model weights and the pretrained weights are placed under .../R3Det_Tensorflow/data/pretrained_weights downloaded from link

Any mistake I am making?

yangxue0827 commented 3 years ago

You need choose the copy corresponding cfgs_res152_dota_r3det_v3.py to cfgs.py

nishit01 commented 3 years ago

ok, I am running this now ... the GPU is not completely getting utilized ... can you help me with that? it seems to me only 1 image is getting processed at a time ... I am not that familiar with TensorFlow, pardon me if I am incorrect?

yangxue0827 commented 3 years ago

use yangxue0827/RotationDetection instead