Detecting Wearable Objects via Transfer Learning
Svatopluk Kraus, Pavel Kršek, Kristína Malinovská, Matúš Tuna · 2019
Transfer learning is a well known technique to circumvent the problem of small datasets in deep machine learning. It has been successfully used in the field of camera surveillance image processing which suffers from poor data quality and quantity. We focused on the task of wearable object detection, namely distinguishing if a person is or is not wearing a backpack. We created new annotations for the DukeMTMC-attribute dataset to overcome the discrepancies among the attributes. We explored transfer learning with a frozen feature extractor as well as the model fine-tuning, which turned out to perform much better. In both setups we found that the Densenet161 is the best from tested architectures. Our best model achieved about 92% balanced accuracy on the testing set.