Sampling Techniques for Large-Scale Object Detection From Sparsely Annotated Objects

Abstract

Efficient and reliable methods for training of object detectors are in higher demand than ever, and more and more data relevant to the field is becoming available. However, large datasets like Open Images Dataset v4 (OID) are sparsely annotated, and some measure must be taken in order to ensure the training of a reliable detector. In order to take the incompleteness of these datasets into account, one possibility is to use pretrained models to detect the presence of the unverified objects. However, the performance of such a strategy depends largely on the power of the pretrained model. In this study, we propose part-aware sampling, a method that uses human intuition for the hierarchical relation between objects. In terse terms, our method works by making assumptions like "a bounding box for a car should contain a bounding box for a tire". We demonstrate the power of our method on OID and compare the performance against a method based on a pretrained model. Our method also won the first and second place on the public and private test sets of the Google AI Open Images Competition 2018.

Solving Missing-Annotation Object Detection with Background Recalibration Loss

“…Training object detectors with incompleteness of annotations is a new rising and challenging vision problem that aims to learn a robust detector by recalibrating the incorrect training signal with partial annotations. In the recent years, a variety of two-stage detector based methods have been proposed, such as part-aware sampling [6] and soft sampling [1]. We observe that two-stage is naturally more robust than the one-stage detectors for the missing-annotation circumstance.…”

Semi-Supervised Object Detection With Sparsely Annotated Dataset

“…Semi-supervised learning for object detection. Object detection using semi-supervised learning is used in situations where it is difficult to manually acquire a sufficient number of annotations to learn, or when pseudo labels are to be obtained from a relatively large number of unlabeled data [16,17,18,19]. In [16], the author's proposed an iterative framework for evaluating and retraining pseudo-labels using pre-trained object detectors and robust trackers to obtain good pseudo-labels in successive frames.…”

“…In [16], the author's proposed an iterative framework for evaluating and retraining pseudo-labels using pre-trained object detectors and robust trackers to obtain good pseudo-labels in successive frames. In [17], it was possible to achieve improved detection performance in the Open Image Dataset V4 by utilizing part-aware sampling and RoI proposals to obtain good pseudo labels for sparsely annotated large-scale datasets. In [18], to efficiently use unlabeled data from the MS-COCO dataset, co-current matrix analysis was used to generate good pseudo labels by using prior information of the labeled dataset.…”

A Robust Training Method for Pathological Cellular Detector via Spatial Loss Calibration

“…They first train the detector using available instance-level annotations, then generate pseudo-annotations, and merge them with the original annotations to iteratively update the detector. For example, Niitani et al ( 22 ) trained the detector to generate annotations using the Open Images Dataset V4 (OID). They then sampled the pseudo-annotations using assumptions such as “cars should contain tires.” However, such a priori assumption in the field of cell detection is unknown.…”

“…Obviously, such an iterative process brings uncontrollability into the training process, e.g., a bad pseudo-annotation generator may significantly influence the final results. In addition, there is not much consensus on how to utilize the pseudo-annotations until now, especially for object detection ( 22 ), e.g., determining the optimal number of iterations is tricky, therefore, it is urgent to solve the SADs training problem in a non-iterative way. Besides, considering that such methods are relatively difficult to replicate, with respect to, empirical and tricky parameter selection or special requirements of the forms of datasets, this study does not include such methods in the comparative experiment.…”