Testing on the training set ? mAp result

[thibo73800]

Hello,

I'm trying to test out your model. I have trained the ssd300 model for 100 epochs. Using your original test script. I get 0.65mAp and the following output :

aeroplane: 1.0 bicycle: 1.0 bird: 0.8 boat: 0.75 bottle: 0.36530612244897953 bus: 0.6666666666666666 car: 0.21311218901580342 cat: 0.2946236559139785 chair: 0.6521739130434783 cow: 1.0 diningtable: 1.0 dog: 0.6451612903225806 horse: 0.6923076923076923 motorbike: 0.4718614718614718 person: 0.08298857633796458 pottedplant: 0.24180327868852458 sheep: 0.6666666666666666 sofa: 0.8 train: 0.7777777777777778 tvmonitor: 1.0 mAP: 0.6560224650525793

But looking at the test script all the images are included rather than only the validation set. After changing the test script to select only the validation set I get the following.

aeroplane: nan bicycle: nan bird: nan boat: 0.125 bottle: nan bus: nan car: 0.22222222222222224 cat: nan chair: 0.0 cow: nan diningtable: nan dog: 0.041666666666666664 horse: 0.16666666666666666 motorbike: 0.125 person: 0.03806079381701687 pottedplant: 0.047619047619047616 sheep: 0.375 sofa: nan train: nan tvmonitor: nan mAP: nan

Even though there is an issue with the mAp computation, most of the class are not detected properly. Do you have any thoughts about the reason for this huge gap in performance?

Thanks,

[ChunML]

That's weird to me. How many images did you run the test on?