Object Detection And Classification On Heterogeneous Datasets

To train an object detection network labeled data is required. More precisely, all objects to be detected must be labeled in the dataset. Here, we investigate how to train an object detection network from multiple heterogeneous datasets to avoid the cost and time intensive task of labeling. In each dataset only a subset of all objects must be labeled. Still, the network shall be able to learn to detect all of the desired objects from the combined datasets. In particular, if the network selects an unlabeled object during training, it should not consider it a negative sample and adapt its weights accordingly. Instead, it should ignore such detections in order to avoid a negative impact on the learning process. We propose a solution for two-stage object detectors like Faster R-CNN (which can probably also be applied to single-stage detectors). If the network detects a class of an unlabeled category in the current training sample it will omit it from the loss-calculation not only in the detection but also in the proposal stage. The results are demonstrated with a modified version of the Faster R-CNN network with Inception-ResNet-v2. We show that the model's average precision significantly exceeds the default object detection performance.