Search code examples
tensorflowobject-detection-api

depth_multiplier and skip_last_stride in Faster RCNN configuration for Tensorflow's Object Detection API

I am going to train 0.5 of Normal Feature Depth using FasterRCNN in Tensorflow.

The FasterRCNN feature extractor has two variables.

depth_multiplier and skip_last_stride.

Do I need to set both as depth_multiplier=0.5 and skip_last_stride=50?

Is this the only place to set?

Solution

  • Yes we just need to update self._depth_multiplier=0.5 at faster_rcnn_mobilenet_v1_feature_extractor.py.

    skip_last_stride and conv_depth_ratio_in_percentage are as they are. Not sure the purpose.

    Then need to update inside

    def _extract_box_classifier_features(self, proposal_feature_maps, scope):

    at depth modification as 

    depth = lambda d: max(int(d * self._depth_multiplier, 16)

    Then all are fine, can go with 0.5 of depth channels for FasterRCNN.