Transfer Learning for new/custom dataset

[soans1994]

Dear Authors,

Great work of pose detection using TensorFlow. I have doubts regarding training a custom dataset. What is the approach for improving the detection of human poses.

1. Should i add the custom dataset annotation json file to the original coco json file and train.
2. Or, is there a way to fine tune the previously trained model for COCO dataset (train from checkpoints or best output model).
3. Or, transfer learn by freezing some layers.

I am not sure the difference between 2 and 3. Can you please help me with some information.

Thank You

[mks0601]

If your custom dataset is similar to MSCOCO, the 1st option is good.
Else if your custom dataset is very small and very different from MSCOCO, then the 2nd option seems good.
The 3rd option seems not good.

[soans1994]

@mks0601 thank you. Yes my dataset is small but contains many images of a person sitting. The coco model doesn’t give good detection because of lack of sitting images from various angles in the coco dataset.
Can you give an example how to train using fine tuning (2nd option), I can train my dataset into a new mode, but I also want the benefit of coco dataset.

[mks0601]

You can just load pre-trained model and continue training on your dataset.

[soans1994]

@mks0601 can you give an example command to continue training in tensorflow. Previously I tried caffe version of openpose for training.