"How about checking your data before going deeper?"
Use TFRecord Viewer to browse contents of TFRecords with object detection/classification annotations.
The viewer runs a Flask server to provide a web gallery with annotation overlays. I.e. you can run it on your server machine, but browse on your local machine.
The web gallery displayed with Fotorama.io.
python3 tfviewer.py datasets/COCO/tfrecord/coco_train.record-00003-of-00100 --labels-to-highlight='car;truck;bus;motorcycle'
python3 tfviewer.py datasets/imagenet/imagenet_fullres/tfrecord/train-00000-of-01024 --overlay classification