Abstract:Solving long-tail large vocabulary object detection with deep learning based models is a challenging and demanding task, which is however <a class="link-external link-http" href="http://under-explored.In" rel="external noopener nofollow">this http URL</a> this work, we provide the first systematic analysis on the underperformance of state-of-the-art models in front of long-tail distribution. We find existing detection methods are unable to model few-shot classes when the dataset is extremely skewed, which can result in classifier imbalance in terms of parameter magnitude. Directly adapting long-tail classification models to detection frameworks can not solve this problem due to the intrinsic difference between detection and <a class="link-external link-http" href="http://classification.In" rel="external noopener nofollow">this http URL</a> this work, we propose a novel balanced group softmax (BAGS) module for balancing the classifiers within the detection frameworks through group-wise training. It implicitly modulates the training process for the head and tail classes and ensures they are both sufficiently trained, without requiring any extra sampling for the instances from the tail <a class="link-external link-http" href="http://classes.Extensive" rel="external noopener nofollow">this http URL</a> experiments on the very recent long-tail large vocabulary object recognition benchmark LVIS show that our proposed BAGS significantly improves the performance of detectors with various backbones and frameworks on both object detection and instance segmentation. It beats all state-of-the-art methods transferred from long-tail image classification and establishes new <a class="link-external link-http" href="http://state-of-the-art.Code" rel="external noopener nofollow">this http URL</a> is available at <a class="link-external link-https" href="https://github.com/FishYuLi/BalancedGroupSoftmax" rel="external noopener nofollow">this https URL</a>.

Long-tail Detection with Effective Class-Margins

Distance Metric-Based Learning for Long-Tail Object Detection

Overcoming Classifier Imbalance for Long-tail Object Detection with Balanced Group Softmax

Margin and Average Precision Loss Calibration for Long-Tail Object Detection

Adaptive Class Suppression Loss for Long-Tail Object Detection

Long-tailed Instance Segmentation using Gumbel Optimized Loss

Rectify the Regression Bias in Long-Tailed Object Detection

Equalized Focal Loss for Dense Long-Tailed Object Detection

The Devil is in Classification: A Simple Framework for Long-tail Object Detection and Instance Segmentation

Balanced Classification: A Unified Framework for Long-Tailed Object Detection

Seesaw Loss for Long-Tailed Instance Segmentation

Classification Calibration for Long-tail Instance Segmentation

Learning from Rich Semantics and Coarse Locations for Long-tailed Object Detection

Class Distribution Aware Focal Loss for Long-Tailed Object Detection

Margin Calibration for Long-Tailed Visual Recognition

Difficulty-aware Balancing Margin Loss for Long-tailed Recognition

Beyond Max-Margin: Class Margin Equilibrium for Few-shot Object Detection

Boosting Long-tailed Object Detection via Step-wise Learning on Smooth-tail Data

Towards Long-Tailed 3D Detection

Adaptive Embedding and Distribution Re-margin for Long-Tail Recognition

The Equalization Losses: Gradient-Driven Training for Long-tailed Object Recognition