PaperlayerView exact revision on arXiv

Source-linked AI summary

Adaptive Class Suppression Loss for Long-Tail Object Detection

Tong Wang, Yousong Zhu, Chaoyang Zhao, Wei Zeng, Jinqiao Wang, Ming Tang

arXiv:2104.00885v1cs.CV

TL;DR

Long-tail object detection methods often group categories by frequency, but suitable partitions vary across datasets. ACSL adapts suppression gradients without frequency-based grouping, and reports improvements on LVIS and Open Images with ResNet50-FPN.

  • Problem

    Long-tail detection has a few frequent categories and many under-represented ones, while frequency-group methods require dataset-dependent category partitions.

    Source: §1 · p. 1, §3.1 · p. 3

  • Method

    ACSL treats categories as scarce and adaptively adjusts suppression gradients based on learning status, without relying on category-frequency grouping.

    Source: §1 · p. 2, §3.1 · p. 4, §3.2 · p. 4

  • Results

    5.18% on LVIS and 5.2% on Open Images are the reported improvements with ResNet50-FPN using ACSL.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Experiments on LVIS and Open Images validate ACSL’s effectiveness, with higher precision especially on tail classes.

    Source: §6 · p. 8

    Group-based methods require dataset-specific frequency partitions; a partition that performs best on LVIS may be suboptimal for Open Images, making selection for new datasets laborious.

    Source: §3.1 · p. 3

Abstract

from arXiv · show

To address the problem of long-tail distribution for the large vocabulary object detection task, existing methods usually divide the whole categories into several groups and treat each group with different strategies. These methods bring the following two problems. One is the training inconsistency between adjacent categories of similar sizes, and the other is that the learned model is lack of discrimination for tail categories which are semantically similar to some of the head categories. In this paper, we devise a novel Adaptive Class Suppression Loss (ACSL) to effectively tackle the above problems and improve the detection performance of tail categories. Specifically, we introduce a statistic-free perspective to analyze the long-tail distribution, breaking the limitation of manual grouping. According to this perspective, our ACSL adjusts the suppression gradients for each sample of each class adaptively, ensuring the training consistency and boosting the discrimination for rare categories. Extensive experiments on long-tail datasets LVIS and Open Images show that the our ACSL achieves 5.18% and 5.2% improvements with ResNet50-FPN, and sets a new state of the art. Code and models are available at https://github.com/CASIA-IVA-Lab/ACSL.

1. Introduction

Long-tail object detection must handle heavily imbalanced category frequencies, while frequency-based grouping can create training inconsistency and weaken discrimination for semantically similar rare categories. ACSL adapts suppression gradients without manual grouping and reports improvements on LVIS and Open Images.

1. Introduction

  • Long-tail datasets have many under-represented rare categories alongside a few frequent categories that contribute most training samples.The paper notes that classic benchmarks tend to have fewer classes and more balanced category distributions.Source: §1 · p. 1
  • Frequency-based grouping can give similarly sized adjacent categories sharply different training strategies and weaken discrimination between semantically similar head and tail categories.The paper illustrates the latter issue with sunglasses and eye mask: ignoring negative gradients from the head category may make the tail classifier confuse them.Source: §1 · p. 2
  • ACSL treats all categories as scarce and adaptively adjusts suppression gradients according to each category’s current learning status.The authors say this avoids manually defining head and tail classes, maintains training consistency, and improves rare-category discrimination.Source: §1 · p. 2
  • The paper presents a statistic-free perspective on long-tail distributions to avoid manual hard division of categories.The proposed ACSL is described as addressing training inconsistency and rare-category discrimination.Source: §1 · p. 2
  • 5.18% on LVIS and 5.2% on Open Images are the reported improvements with ResNet50-FPN using ACSL.The authors report these results as validation of ACSL’s effectiveness.Source: §1 · p. 2

2. Related Works

Prior work addresses long-tail recognition through sampling, reweighting, specialized losses, or changes to classifier training. These approaches often rely on category-frequency information, which ACSL is designed to avoid.

2. Related Works

  • Re-sampling and re-weighting seek to rebalance training, while class-balanced loss uses effective sample numbers to reweight the loss.Class-aware sampling and repeat-factor sampling vary sampling frequencies by category; re-weighting emphasizes tail samples.Source: §2 · p. 3
  • Specialized losses and training strategies include ignoring head-class negative gradients, grouping categories for within-group softmax, and retraining a balanced classifier.Other cited approaches use a label-distribution-aware margin, an unbiased softmax extension, or bi-level class-balanced sampling.Source: §2 · p. 3
  • The surveyed approaches can relieve imbalanced training, but they depend on class-distribution information and may be inconvenient to apply to new long-tail datasets.The paper positions its loss as not relying on category-frequency priors.Source: §2 · p. 3

3. Adaptive Class Suppression Loss

ACSL replaces hard head/tail grouping with confidence-driven, per-sample suppression, aiming to preserve useful discrimination while avoiding excessive negative gradients on rare classes.

3.1. Limitations of Group-based Methods

  • Group-based methods require dataset-tuned cutoffs: LVIS performs satisfactorily across 50–500, while a partition suited to LVIS may not suit Open Images.The table compares grouping strategies; the figure plots sorted category index against the base-10 logarithm of instance count.Source: Table 1 · p. 3, §3.1 · p. 3, Figure 2 · p. 4

3.2. Formulation of ACSL

  • For each sample, ACSL suppresses non-target classes only when their confidence exceeds ξ, preserving negatives for confusing classes while avoiding unnecessary suppression.The method uses output confidence rather than category-frequency statistics to choose which categories to suppress.Source: §3.2 · pp. 4–5

3.3. Advantages over Previous Methods

  • ACSL sets suppression per sample from current confidences rather than fixed class statistics, and works on LVIS and Open Images with one hyper-parameter setting.Earlier approaches use label distribution for their weights or category groups, whereas ACSL can assign different category weights to samples from the same class.Source: §3.3 · p. 5

4. Experiments on LVIS

The LVIS experiments evaluate ACSL across training choices, backbones, detector types, and category frequencies. They report stronger comparisons with prior methods and the largest gains on rare categories.

4.1. Dataset and Setup

  • LVIS-v0.5 has 1,230 categories; the experiments use 57k training images and 5k validation images, with frequent, common, and rare categories evaluated separately.The frequency groups are frequent (>100 images), common (11–100), and rare (<10); the reported metrics include mAP, APr, APc, and APf.Source: §4.1 · pp. 5–6

4.5. Performance Analysis

  • ACSL’s precision curves remain close to baseline for frequent classes, improve moderately for common classes, and diverge most for rare classes.Figure 4 plots precision against sorted class index for each frequency group.Source: Figure 4 · p. 6, §4.5 · p. 7

4.2. Implementation Details

  • The two-stage setup trains a naive Faster R-CNN detector first, then fine-tunes its classifier with specialized losses and subsamples background negatives by category frequency.Given an approximate rare:common:frequent image ratio of 1:10:100, the method uses 1% of background samples for rare classes and 10% for common classes.Source: §4.2 · p. 6

4.3. Ablation Study

  • Even after 24 training epochs, the baseline’s rare-category accuracy is 4.3%; more iterations improve rare and common categories and raise overall mAP by 1.1%.The authors identify tail classes as the baseline’s performance bottleneck.Source: §4.3 · p. 6
  • With ξ = 0.7, ACSL reaches 26.36% overall mAP, 5.18% above baseline; further increasing ξ deteriorates performance by weakening classifier discrimination.The authors report that increasing ξ initially relieves suppression on tail classifiers, while very small values suppress too many categories.Source: §4.3 · p. 6

4.6. Comparison with State of the Arts

  • 26.36% mAP: ACSL surpasses Equalization Loss (25.06% mAP) and BAGS (25.96% mAP), with the best single ResNet-50 results on rare and common categories.The reported rare-category AP is 18.64%, and the authors state that head-category precision is not sacrificed.Source: Table 4 · p. 7, §4.6 · pp. 7–8

4.4. Generalization on Stronger Models

  • With stronger backbones, ACSL exceeds baseline by 5.13% mAP with ResNet101 and 4.23% with ResNeXt-101; it also exceeds Cascade R-CNN baselines.For ResNet101, the reported rare-category APr improvement is 16.11%; with ResNet-101 Cascade R-CNN, ACSL reaches 29.71% mAP, 4.57% above baseline.Source: §4.4 · p. 7

4.6. Comparison with State of the Arts

  • 29.47% mAP: ACSL reaches this best reported score with multi-scale testing; it also outperforms the listed methods with ResNet101 and ResNeXt-101-64x4d backbones.The reported ResNet101 score exceeds Equalization Loss and BAGS by 1.44% and 1.1% mAP, respectively; the ResNeXt-101-64x4d score is 28.93% mAP.Source: Table 4 · p. 7, §4.6 · p. 8

5. Experiments on Open Images

On Open Images, ACSL outperforms baseline models across the reported backbones, with gains also reported for rare categories. The experiments use the dataset’s validation split and compare methods with the same backbone.

5. Experiments on Open Images

  • 60.3% AP: ACSL exceeds the ResNet50-FPN baseline by 5.2%; with ResNet152, it reaches 62.8% AP, 5.4% above baseline.The experiments use a multiple-label setting and set ξ to 0.7.Source: §5 · p. 8
  • For rare category “Face powder” (“Fac” for short), ACSL’s performance gain is 63.1% AP; it also brings considerable improvements to other rare categories, leading to higher overall precision.Source: §5 · p. 8
  • The same-backbone comparison with other Open Images methods reports a performance advantage for ACSL.Table 7 covers ResNet50-FPN models evaluated on 500 categories.Source: Table 7 · p. 8, §5 · p. 8

6. Conclusion

The conclusion presents ACSL as a loss for long-tail datasets that reduces over-suppression of tail classifiers while preserving discrimination between easily confused categories. Experiments on LVIS and Open Images validate its effectiveness, and extending it to long-tail image classification is future work.

  • ACSL is designed to prevent head-category samples from over-suppressing tail classifiers while preserving discrimination between easily confused categories.Source: §6 · p. 8
  • Experiments on LVIS and Open Images validate ACSL’s effectiveness, with higher precision reported especially for tail classes.Source: §6 · p. 8
  • Applying ACSL to long-tail image classification datasets is identified as future work.Source: §6 · p. 8

Related papers

Loading 2104.00885v1…