ucbdrive / few-shot-object-detection

Implementations of few-shot object detection benchmarks
Apache License 2.0
1.08k stars 225 forks source link

Why is the mAP50 on Vocsplit1 so low? #161

Open HuangLian126 opened 2 years ago

HuangLian126 commented 2 years ago

I use the Base Model for Split1 provided in the link. I fine-tuning the model for 1-shot setting. However, the mAP50 on Voc split1 is so low.

mAP50: aeroplane bicycle boat bottle car cat chair diningtable dog horse person pottedplant sheep train tvmonitor
bird( 3.328) bus( 62.504) cow(63.879) motorbike(10.217) sofa (10.217)

87.043, 86.007, 69.354, 73.429, 79.536 , 87.813, 51.231 , 56.622 , 83.109 , 86.443 , 61.499 , 42.576 , 72.421 , 87.628 , 75.419 3.328 ,62.504 , 63.879 , 10.217 , 4.595

My hyperparameters: gpu: 1 batchsize: 4 lr: 0.000125 step: 12000 max_iter: 16000