hello guys. I’m trying to use tensorflow object detection API’s models on my data. I trained faster rcnn and retinanet on my data, however after 100.000 steps on about 3000 images I didn’t get any good results. and mAP is about 0.36!!
I have no idea what should I do?
Also:
as my data is based on video series, I split the data based on each video and use frames as input.
and I had work on other data, and I noticed that with google colab V100 or A100 GPU’s, there is no need to change the learning rate, but on my GPU I must divide original learning rate to the new batch size. so in this project I did the same.
I used some appropriate augmentation too, but result hadn’t changed!
my errors in loss graphs are low. both in training and eval, so there is no ovrfiting, I think so!
could you please give me some tips to improve my results.
Thank you in advance.
With the details provided above, Bad results may be due to your low in traing data size, class imbalance in the data,you’ve tried Faster R-CNN and RetinaNet- these models architectures may not performing better on your specific dataset.May be the models are underfitting with current data you have.
You can try following steps , it might help in improving mAP.
1.Try different model architectures(Like YOLO, EfficientDet, or Mask R-CNN etc…)
2. Try different learning rate schedules, optimizer types, and regularization techniques (e.g., dropout, weight decay).
3. Try training for more steps.
4.Try some effective augmentations include random cropping, flipping, rotation, scaling, and color jittering.
Try to check this colab tutorial from model garden object detection where detailed instructions are given for fintuning.
Hello and thanks for your responds. I added more data also I tried different learning rates and data augmentation but all runs follow same patterns. I have a huge gap between train and eval.
as I’m using tensorflow object detection API, is it possible to change the optimizer? because I tried to add other augmentations and change the regularization but it didn’t work!
now I decide to use keras instead of API, I think I can modify models in a better way.
