Class weights for balancing data in TensorFlow Object Detection API

Question

I\'m fine-tuning SSD object detector using TensorFlow object detection API on Open Images Dataset. My training data contains imbalanced classes, e.g.

1. top (5K i

Answer 1

The Object Detection API losses are defined in: https://github.com/tensorflow/models/blob/master/research/object_detection/core/losses.py

In particular, the following loss classes have been implemented:

Classification losses:

1. WeightedSigmoidClassificationLoss
2. SigmoidFocalClassificationLoss
3. WeightedSoftmaxClassificationLoss
4. WeightedSoftmaxClassificationAgainstLogitsLoss
5. BootstrappedSigmoidClassificationLoss

Localization losses:

1. WeightedL2LocalizationLoss
2. WeightedSmoothL1LocalizationLoss
3. WeightedIOULocalizationLoss

The weight parameters are used to balance anchors (prior boxes) and are of size [batch_size, num_anchors] in addition to hard negative mining. Alternatively, the focal loss down weighs well classified examples and focusses on the hard examples.

The primary class imbalance is due to many more negative examples (bounding boxes without objects of interest) in comparison to very few positive examples (bounding boxes with object classes). That seems to be the reason why class imbalance within positive examples (i.e. unequal distribution of positive class labels) is not implemented as part of object detection losses.