Which datasets have the Fast R-CNN detectors been trained on?

[tlind]

Hi, since training data matters a lot when comparing different methods, could you clarify which datasets the Fast R-CNN detectors in the two provided example models (RGB and DepthJet) have been pre-trained on?

In the Adaptive Fusion paper by Mees et al., it is stated that the RGB network has been fine-tuned using a model from the Caffe model zoo. Was this a model trained on COCO (+ the usual ImageNet pretraining)? Or were you just copying the weights of the feature extractor trained on ImageNet, since for the 5-class use-case, the classification/detection layers anyway had to be retrained from scratch?

For the DepthJet network, is it trained completely from scratch on the hospital dataset? Or was any pretraining on e.g. ImageNet or COCO involved?

Thanks!

[marinaKollmitz]

Hi, for both networks, we initialized the convolutional layers with the pretrained model from here and only finetuned the fully connected layers on the RGB/DepthJet images. The train set of the RGB model also included some images from the InOutDoorPeople dataset from the adaptive fusion paper. I am planning to upload the training code to github very soon.