Trying to reproduce paper results HRSC2016

[Artcs1]

Hi, First Great work.

I am trying to reproduce HRSC results as the KLD paper, but I always get like 0.03 AP below for each method. I am using bs 4, and training for 20 epochs with the cfgs in the corresponding 'configs' folder. Do you have any clue what is going on?

Greetings

[yangxue0827]

Try bs=1 and use python train.py

[Artcs1]

I run a new experiment as suggested but I am still having a lower AP than reported. Just to clarify in configs file (KL_FUNC = sqrt, KL_TAU = 2.0). Were those values ​​used to generate the results of the original paper?

[yangxue0827]

how about this config: https://raw.githubusercontent.com/yangxue0827/RotationDetection/main/configs/HRSC2016/kl/cfgs_res50_hrsc2016_kl_v1.py do you reproduce the performance？

[Artcs1]

I am using that configuration with a RTX A4000: My output was:

rotation eval: Writing ship VOC resutls file Threshold: 0.5 cls : ship|| Recall: 0.9332247557003257 || Precison: 0.30285412262156447|| AP: 0.8559664693973696 F1:0.8593960007045868 P:0.8784013605442177 R:0.8412052117263844 mAP is : 0.8559664693973696

Threshold: 0.55 cls : ship|| Recall: 0.9210097719869706 || Precison: 0.29889006342494717|| AP: 0.8488067570226043 F1:0.8544043201723106 P:0.8732993197278912 R:0.8363192182410424 mAP is : 0.8488067570226043

Threshold: 0.6000000000000001 cls : ship|| Recall: 0.9014657980456026 || Precison: 0.2925475687103594|| AP: 0.8227167138601044 F1:0.845252905863143 P:0.8639455782312925 R:0.8273615635179153 mAP is : 0.8227167138601044

Threshold: 0.6500000000000001 cls : ship|| Recall: 0.8745928338762216 || Precison: 0.28382663847780126|| AP: 0.7719636329409395 F1:0.8302778642663381 P:0.8486394557823129 R:0.8127035830618893 mAP is : 0.7719636329409395

Threshold: 0.7000000000000002 cls : ship|| Recall: 0.8257328990228013 || Precison: 0.2679704016913319|| AP: 0.7521026220294736 F1:0.798663887562038 P:0.8163265306122449 R:0.7817589576547231 mAP is : 0.7521026220294736

Threshold: 0.7500000000000002 cls : ship|| Recall: 0.74185667752443 || Precison: 0.24075052854122622|| AP: 0.6430973163191885 F1:0.7287434102205883 P:0.766397124887691 R:0.6946254071661238 mAP is : 0.6430973163191885

Threshold: 0.8000000000000003 cls : ship|| Recall: 0.5773615635179153 || Precison: 0.1873678646934461|| AP: 0.4266389646833525 F1:0.5775259817801465 P:0.6073674752920036 R:0.5504885993485342 mAP is : 0.4266389646833525

Threshold: 0.8500000000000003 cls : ship|| Recall: 0.30700325732899025 || Precison: 0.09963002114164905|| AP: 0.19439607500936473 F1:0.30941593507136134 P:0.3296500920810313 R:0.2915309446254072 mAP is : 0.19439607500936473

Threshold: 0.9000000000000004 cls : ship|| Recall: 0.06921824104234528 || Precison: 0.022463002114164906|| AP: 0.09090909090909091 F1:0.07111889102245027 P:0.0851305334846765 R:0.061074918566775244 mAP is : 0.09090909090909091

Threshold: 0.9500000000000004 cls : ship|| Recall: 0.003257328990228013 || Precison: 0.0010570824524312897|| AP: 0.0036363636363636364 F1:0.0034522429356127087 P:0.003683241252302026 R:0.003257328990228013 mAP is : 0.0036363636363636364

mAP50:95 : 0.5410234005807852