Not All Labels Are Equal - Rationalizing The Labeling Costs for Training Object Detection
Created: =dateformat(this.file.ctime,"dd MMM yyyy, hh:mm a") | Modified: =dateformat(this.file.mtime,"dd MMM yyyy, hh:mm a")
Tags:
Annotations
Recently, several methods have been adapted specificallyfor the task of object detection [6, 7, 19–22], some of whichare based on the core-set approaches where the diversityof the training examples is taken into account. However,the state-of-the-art approaches are based on the uncertainty[4, 5, 8, 23]. The work of [23] consists of an ensemble ofobject detectors that provide bounding boxes and probabil-ities for each class of interest. Then, a scoring function isused to obtain a single value representing the informative-ness of each unlabeled image. Similar to that is the workof [8] where the authors compute the instance-based uncer-tainty. Another work [4] gives an elegant solution, reachingpromising results compared with other single-model meth-ods. The authors train a network in the task of detectionwhile learning to predict the final loss. In the sample ac-quisition stage, samples with the highest prediction loss areconsidered the most interesting ones and are chosen to belabeled. In the state-of-the-art approach [5], authors definethe aleatoric and epistemic uncertainty, in both class andbounding box level, and use the combined score to deter-mine the images that need labeling. Our work is related butdifferent from the above-mentioned works. Similarly, weconsider the uncertainty of the detector as part of the solu-tion. Unlike them, we find that the robustness of the detectoris even more reliable as an acquisition function, especiallyfor the low-performing classes. We then unify these twoscores to reach high performance in the majority of classes. * show annotation
A predic-tion that has a high entropy suggests that the object is highlydissimilar to the images the network is trained on. Thus, iflabeled, it will provide different information to the ones wehave. * show annotation
only an un-certainty-based acquisition function is not an ideal solution,especially for images coming from low-performing classes. * show annotation
Thehigher the inconsistency, the more informative the sampleis for training and therefore potentially worth labeling. * show annotation
ntuition behind using the maximum score instead ofsome other score, such as the average, is that labeling an * show annotation
mage that has at least one difficult object, independentlyof the number of easy objects is beneficial because of thedifficult object. * show annotation