Training Region-based Object Detectors with Online Hard Example Mining (1604.03540v1)

Abstract: The field of object detection has made significant advances riding on the wave of region-based ConvNets, but their training procedure still includes many heuristics and hyperparameters that are costly to tune. We present a simple yet surprisingly effective online hard example mining (OHEM) algorithm for training region-based ConvNet detectors. Our motivation is the same as it has always been -- detection datasets contain an overwhelming number of easy examples and a small number of hard examples. Automatic selection of these hard examples can make training more effective and efficient. OHEM is a simple and intuitive algorithm that eliminates several heuristics and hyperparameters in common use. But more importantly, it yields consistent and significant boosts in detection performance on benchmarks like PASCAL VOC 2007 and 2012. Its effectiveness increases as datasets become larger and more difficult, as demonstrated by the results on the MS COCO dataset. Moreover, combined with complementary advances in the field, OHEM leads to state-of-the-art results of 78.9% and 76.3% mAP on PASCAL VOC 2007 and 2012 respectively.

• The paper introduces OHEM to automatically select challenging samples during training, eliminating complex heuristics and extensive hyperparameter tuning.
• The method integrates hard example mining into the SGD process, leading to faster convergence and lower training loss by focusing on high-loss regions.
• Empirical results reveal significant mAP improvements on benchmarks like PASCAL VOC and MS COCO, underscoring the method’s scalability and efficiency.

Training Region-based Object Detectors with Online Hard Example Mining

The paper "Training Region-based Object Detectors with Online Hard Example Mining" addresses a significant challenge in the field of object detection, particularly in the context of training deep convolutional neural networks (ConvNets). The authors, Abhinav Shrivastava, Abhinav Gupta, and Ross Girshick, propose an Online Hard Example Mining (OHEM) algorithm to streamline and enhance the training process for these models.

Overview

Region-based object detectors have achieved considerable success, yet the training procedures remain bogged down by numerous heuristics and hyperparameters. These add complexity and necessitate extensive hyperparameter tuning. The primary contribution of this paper is the introduction of OHEM, a method designed to automatically select hard examples during training, thus eliminating the need for several heuristics and hyperparameters.

The genesis of OHEM lies in the recognition that object detection datasets are inherently imbalanced, containing a vast majority of easy examples compared to a small subset of difficult ones. Traditional hard example mining techniques, such as those used in deformable part models (DPM) and shallow neural networks, rely on offline alternation between model updates and example selection. OHEM adapts this paradigm to the online learning environment of modern ConvNet-based detectors, making it possible to train these models more efficiently.

Technical Contributions

The OHEM algorithm modifies the Fast R-CNN training approach by integrating a mechanism for hard example selection during the stochastic gradient descent (SGD) process. Unlike previous methods that freeze the model periodically to find new hard examples, OHEM evaluates all region of interest (RoI) proposals in a single forward pass. It then selects hard examples based on their loss values and performs backpropagation exclusively on this subset, thus maintaining the efficiency of learning.

Key aspects of OHEM include:

1. Elimination of bg_lo Heuristic: Traditional Fast R-CNN uses a bg_lo threshold to balance foreground and background examples. OHEM removes this heuristic by allowing the model to focus on examples with the highest loss, leading to unbiased and more effective example selection.
2. Improved Optimization: By focusing on hard examples, OHEM achieves lower training loss and faster convergence. This method effectively reduces the noise introduced by easy examples which contribute little to the learning process.
3. Scalability: The benefits of OHEM are amplified as the dataset size increases, making it particularly effective on large and challenging datasets like MS COCO.

Empirical Results

The authors validate their approach through comprehensive experiments on multiple benchmarks:

• PASCAL VOC 2007: OHEM improves mean Average Precision (mAP) significantly, from 67.2% to 69.9% using VGG16 networks.
• PASCAL VOC 2012: OHEM achieves a gain of 4.1 percentage points in mAP, underscoring its effectiveness.
• MS COCO: The method demonstrates substantial improvements across various evaluation metrics. For instance, OHEM increases AP from 19.7% to 22.6% and achieves a marked increase in AP for medium-sized objects.

Implications and Future Work

The OHEM method presents a practical improvement for training region-based object detectors, offering both theoretical and empirical advancements:

• Practical Implications: The removal of heuristics streamlines the training process, allowing for more robust and scalable model development. This contributes to faster and more efficient deployment of object detection systems in real-world scenarios.
• Theoretical Implications: The approach challenges the necessity of intricate heuristics in ConvNet training, advocating for data-driven example selection mechanisms.

Future developments may explore further integration of OHEM with other advanced techniques in deep learning, such as multi-scale training and iterative bounding-box regression. Additionally, expanding the use of OHEM in different network architectures and tasks within computer vision could provide broader insights into its applicability.

In conclusion, the paper presents a well-founded method that enhances the training of region-based ConvNets by leveraging online hard example mining. This advance not only elevates detection performance but also simplifies the training process, making it a valuable contribution to the object detection community.