search

VDIGPKU / M2Det

M2Det: A Single-Shot Object Detector based on Multi-Level Feature Pyramid Network
MIT License
1.45k stars 318 forks source link

why i cannot get a good resuilt? #92

Open maxiaoming6669 opened 5 years ago

maxiaoming6669 commented 5 years ago

I trained 15,000 data and applied the pre-training model. The lowest LOSS returned during training was only 0.5. None of the test results were detected, even in the test training set. I use VOC-type datasets

williamluke4 commented 5 years ago

Yeah, I'm having the same Issue, I trained on COCO 2014. image

maxiaoming6669 commented 5 years ago

Who can tell me why?

hshsjkjy commented 5 years ago

Has anyone trained a good result? The two loss values have been around 1 and the detection is very poor.

kuwt commented 5 years ago

Yeah, I'm having the same Issue, I trained on COCO 2014. image

I got the same result. image

nile649 commented 5 years ago

Same here

williamluke4 commented 5 years ago

This may shed some light https://github.com/qijiezhao/M2Det/issues/81#issuecomment-510857633 Translated "The effect of training is not good, of course, it must be solved from training. Adjusting the hyperparameters such as lr batch during training and training the model well, this problem is naturally solved. If the training effect is not good, or depends on how your current specific hyperparameters are set, you can record the training loss and test accuracy during training, and then adjust the hyperparameter according to the loss curve accuracy curve."

shuoshuoa commented 5 years ago

i also has bad result in KITTI.

zhangxiang1209 commented 5 years ago

I get the same issue, why??? how to solve it??? image