How much impact the trainval's DATASET.MGDPT_IMG_RESIZE have

[FujiwaraZayako]

Due to equipment problems, during the experiment，my trainval's DATASET.MGDPT_IMG_RESIZE, once set to 640, the GPU video memory will be insufficient and I have to reduce it. I want to know how much this parameter affects the final training results

[Coronal-Halo]

I think it depends on your application, but I feel that if you compress the image pairs in any way, it would negatively affect the performance of the model. I tested the model for pose estimation and for my specific case I find using resized (smaller) images decreases my accuracy. So, if your original images are bigger than 640 x 640, it will affect your training results.

[FujiwaraZayako]

I've tried entering 560size before。 From the perspective of images, the input size is bound to affect the relevant information and attention coding of the detected image pair. Smaller input will indeed reduce the accuracy. However, from my model, it seems that the effect is not that big.In contrast, the impact of sample size and learning strategies is much larger.