Atkārtot eksperimentus ar lielāku resolution un batch size
Ievākt datus arī par AP75 AP50 (jēgpilnāk var saprast kā modelis darbojas)
Aprakstīt SLR daļu pēc šī piemēra overleafā: http://share.yellowrobot.xyz/quick/2024-2-9-3F192320-2147-4C88-BC63-A583D67003F3.pdf
Small model 700x400, 0.14 boxAP
State of art 0.44 boxAP
AP (Average Precision) came from COCO challange with pre-defined step thresholds.
AP@[0.5:0.95] = vidējais starp AP pie IoU thresholdiem 0.5, 0.55 … 0.95
Higher AP in this range does not mean that the model is better, it just means that the model is better at detecting objects at different IoU thresholds. For example, a model with AP@[0.5:0.95] = 0.3 and AP@[0.5] = 0.9 is better than a model with AP@[0.5:0.95] = 0.2 and AP@[0.5] = 0.9, because the first model is better at detecting objects at different IoU thresholds, even though the second model is better at detecting objects at IoU threshold 0.5.
Thus AP is strange metric, because it is not clear what is the best value for it. It is better to use AP@[0.5] and AP@[0.75] to compare models.
For training COCO most use Object detection package
https://mmdetection.readthedocs.io/en/latest/overview.html
https://openmmlab.com/ecosystem
If for loss plot using batches instead of epochs then need to calculate how many batches are in one epoch. For example, if we have 140000 samples and batch size is 8 then we have 140000/8=17500 batches in one epoch.