This repo is a forked version and aim to implement the method proposed in Omnia faster r-cnn. This paper introduces an offline method for generating pseudo-label instances in a merged dataset which can contain many missing-label instances. For a brief introduction on merged datasets, interested readers are referred to here.

Instead of faster r-CNN, we used YOLO v3 (implemented). First, two disjoint datasets are created, using VOC family of datasets, as follows:

• VOC 2007; with following focused categories: ={cat, cow, dog, horse, train, sheep}, the datatset is called VOC7_A.
• VOC 2012; ={car, motorcycle, bicycle, aeroplane, bus, person}, the dataset is named VOC12_B.

Using function incld_or_excld_dataset in utils/custom_datasets.py, we load VOC7_A (VOC12_B) from VOC2007 (VOC2012). The categories of interest for VOC7_A are indicated in voc7-voc12-Exp1/data/voc2007.txt.

Then, two YOLOs are trained separately on each of the above datasets to be used later for generating pseudo labels for missing label instances from either categories or .

More precisely, after merging together VOC7_A and VOC12_B, the resultant dataset can contain missing label instances. For example, in the following annotated image from VOC7_A, it contains horse annotated, but no annotation for person as "person" category is not in . However, after merging VOC7_A and VOC12_B, the person in this image becomes a missing label instance as "person" is an object of interest in .

Figure 1: A sample from VOC7_A, which contains the annotated instances from {cat, cow, dog, horse, train, sheep}. While the "Person" instance is not an object of interest in VOC7_A, it becomes a missing-label instance after merging VOC7_A and VOC12_B.

In conclusion, using YOLO trained on VOC12_B, the authors proposed to generate pseudo label for the possible missing-label instances from in VOC7_A. Simialrly, using YOLO trained on VOC7_A, the missing label instance that exist in VOC7_B can be generated.

As main point of start, in order to train a YOLO model, test it, or generate pseudo_label for missing label instance, offline_ODs.py should be used.

required packages:

• matplotlib
• opencv
• tqdm
• torchvision
• pillow
• torch

Reference

Rame, E. Garreau, H. Ben-Younes, and C. Ollion, “Omnia faster r-cnn: Detection in the wild through dataset merging and soft distil-lation,”arXiv preprint arXiv:1812.02611, 2018.