PaperlayerView exact revision on arXiv

Source-linked AI summary

To Boost or Not to Boost? On the Limits of Boosted Trees for Object Detection

Eshed Ohn-Bar, Mohan M. Trivedi

arXiv:1701.01692v1cs.CV

TL;DR

The paper examines whether boosted decision trees can exploit greater model capacity and larger datasets for object detection. It studies augmentation and training choices, reports strong Caltech detection results, and concludes that increasing tree capacity has limited benefits.

  • Problem

    The limits of the relationship between boosted-tree capacity and dataset size for object detection had not been fully explored, including whether performance had saturated.

    Source: §I · pp. 1–2

  • Method

    The study varies dataset size and properties alongside decision-tree depth, using augmentation experiments to investigate boosted detectors.

    Source: §I · p. 2, A. Boosting Framework · p. 2

  • Results

    18.29% miss rate on the MRN-4 metric was reported as the best detection result to date among CNN and non-CNN methods on the improved-annotation Caltech dataset.

    Source: B. Results on Caltech-New · p. 5

  • Takeaways & Limitations

    Increasing weak-learner complexity saturates early, and further tree-depth increases provide limited classification-power gains.

    Source: §III · p. 5, §IV · p. 6

    Boosted trees show limited representation ability: performance saturates at depth 5/6 even with extensive augmentation.

    Source: §III · p. 5, §IV · p. 6

Abstract

from arXiv · show

We aim to study the modeling limitations of the commonly employed boosted decision trees classifier. Inspired by the success of large, data-hungry visual recognition models (e.g. deep convolutional neural networks), this paper focuses on the relationship between modeling capacity of the weak learners, dataset size, and dataset properties. A set of novel experiments on the Caltech Pedestrian Detection benchmark results in the best known performance among non-CNN techniques while operating at fast run-time speed. Furthermore, the performance is on par with deep architectures (9.71% log-average miss rate), while using only HOG+LUV channels as features. The conclusions from this study are shown to generalize over different object detection domains as demonstrated on the FDDB face detection benchmark (93.37% accuracy). Despite the impressive performance, this study reveals the limited modeling capacity of the common boosted trees model, motivating a need for architectural changes in order to compete with multi-level and very deep architectures.

I. INTRODUCTION

The paper studies whether boosted decision trees can use increased model capacity and dataset size to improve object detection. Its experiments reveal limits to that approach while also producing stronger ACF and LDCF detectors.

I. INTRODUCTION

  • The study varies augmentation-driven dataset properties and tree depth to examine boosted decision trees, then applies its findings to pedestrian and face detection.The unresolved question is whether tree capacity has saturated or can still benefit from additional data.Source: Figure 1 · p. 1, §I · p. 2
  • The modified ACF and LDCF detectors improve over their baselines, with LDCF84++ reported as the best-performing non-CNN Caltech detector to date.The paper reports about 10% points of improvement over ACF while retaining 10 HOG+LUV channels.Source: Figure 2 · p. 1, §I · p. 2
  • The relationship between tree capacity and dataset size had not been fully explored for detection, leaving its saturation point uncertain.The authors report limited generalization from increasing weak-learner capacity, with performance saturating quickly despite added augmentation.Source: §I · pp. 1–2
  • The experiments identify dataset properties and training practices that support consistent detection gains, while finding video-based augmentation limited against other options.The study avoids deep CNN features to isolate the strengths and limitations of boosting with simple channels.Source: §I · p. 2

A. Boosting Framework

The study uses a soft cascade of decision-tree weak learners and examines how to train boosted models efficiently as dataset size and tree depth increase.

A. Boosting Framework

  • The detector uses a soft cascade with decision trees as weak learners, whose non-leaf nodes make binary feature-threshold decisions.Maximum depth is an important parameter; prior detection work generally used depth 2, while recent studies proposed depths 3–4 for larger datasets.Source: A. Boosting Framework · p. 2
  • Efficient training is a central concern because stump training costs O(K × N) for K features and N samples.The paper notes that fast training matters for algorithm development and comparison with CNN detectors trained on millions of samples.Source: A. Boosting Framework · p. 2
  • The adopted quick-boosting procedure quantizes feature values into 256 bins and samples a subset of features during search.The authors report that the common 1/16 feature-sampling strategy can vary in performance by 3–4% across runs.Source: A. Boosting Framework · p. 2
  • Exhaustively searching the first 512 weak learners in the final hard-negative stage, then returning to 1/16 sampling, reduces run-to-run variability to about 1%.The authors use this compromise to balance reproducibility with speed rather than performing exhaustive search in every iteration.Source: A. Boosting Framework · p. 2

B. Augmentation Techniques

The paper examines video sampling and several image augmentations as ways to increase training data for higher-capacity boosted detectors.

B. Augmentation Techniques

  • Dense video sampling increases dataset size, but the authors identify issues that can produce sub-optimal detector performance.They frame augmentation choices as important because additional data regularizes training for models with increased capacity.Source: B. Augmentation Techniques · p. 2
  • The studied augmentations include PCA-based color alteration, horizontal flipping, and translations controlled by maximum shift and sampling interval.For example, T31 generates four additional samples shifted one pixel west, east, north, or south.Source: B. Augmentation Techniques · p. 3
  • Scale+Crop scales and recenters each positive sample before cropping, preserving the box center while jittering its width or height.The authors propose it after observing classifier sensitivity to small noise in ground-truth location.Source: B. Augmentation Techniques · p. 3
  • The study also considers adding more heavily occluded pedestrian samples and transferring foreground/background using Caltech visibility masks.These approaches are presented as ways to enlarge the dataset and examine generalization.Source: B. Augmentation Techniques · p. 3

C. Model Settings

The experiments use a fixed pedestrian detection window and compare several Caltech video-sampling rates.

C. Model Settings

  • The model is 41 × 100 pixels, and padding makes the sliding window 64 × 128 pixels.The images are upsampled by one octave to handle pedestrians at least 50 pixels tall in the reasonable test setting.Source: C. Model Settings · p. 3
  • Most augmentation experiments use Caltech×3, sampling every 10th video frame.The paper also studies Caltech×7.5, Caltech×15, and Caltech×30, with the latter sampling every frame.Source: C. Model Settings · p. 3
  • The additional sampling settings include every 4th frame for Caltech×7.5 and every-frame sampling for Caltech×30.Caltech×15 is also included among the higher-sampling conditions.Source: C. Model Settings · p. 3

III. EXPERIMENTAL ANALYSIS

The Caltech experiments examine how training data and augmentation interact with boosted-tree capacity in pedestrian detection. Performance improves substantially, but the results also expose limits in the augmentation the classifier can use and in gains from deeper trees.

Flip augmentation

  • Removing flips from the L3N50 baseline reduces miss rate from 29.28% to 26.67%, while the two revised baselines perform similarly and the no-flip version trains faster.Aspect-ratio standardization produces non-aligned boxes; the paper attributes the flip issue to adding more poorly aligned samples.Source: §III · p. 3

Additional augmentation

  • L3N50 variants benefit from scaling plus crop, best with a 1.1 scaling factor, and 1-pixel translation, but not color augmentation.Increasing negatives to L3N100 did not further improve performance, so the authors increased model capacity.Source: §III · p. 3

Increased model capacity

  • 20.85% miss rate is reached with the fast ACF model on Caltech×3 after increasing capacity to L5N50; scaling helps more than translation.Even slight off-center shifts hinder performance, and the classifier does not handle them well.Source: §III · pp. 3–4

Video augmentation

  • Caltech×3 required less video augmentation than Caltech×10 to reach best performance; denser sampling needed 200k negatives for comparable performance, and further augmentation or capacity did not beat 20.69%.The paper reports no visible improvement from further increasing data and says tree depth beyond 6 leads to overfitting.Source: §III · p. 4

Better features

  • 20.69% miss rate is reported for ACF, nearly matching Checkerboards at 18.47%; LDCF84 reaches 17.15%, and without feature-pyramid approximation it reaches 15.40%.The approximated LDCF84 model runs 5 times faster than Checkerboards.Source: §III · p. 4

Summary

  • The boosted detector handles only mild deformations: color jitter is unhelpful, flips are problematic, and performance saturates at depth 5/6 even with extensive augmentation.Further tree depth offers limited classification-power benefit compared with adding layers in CNN architectures.Source: §III · pp. 4–5

Run-time analysis

  • ACF-L5 and LDCF84 run at 6.7 and 2.5 fps, respectively, faster than top feature-rich models at 0.1–1 fps.Checkerboards runs at 0.5 fps.Source: §III · p. 5

A. Context Analysis

The context analysis finds that adding contextual reasoning improves both ACF and LDCF84 models, while increasing padding alone does not; a location-prior approach offers further improvement.

A. Context Analysis

  • Increasing model padding to represent context did not affect performance, whereas a second approach improved it further.The authors state that the improvement hints at remaining room for context modeling.Source: A. Context Analysis · p. 5
  • The location-prior approach forms features from products of position, size, and confidence, then recomputes object scores with an SVM.The resulting feature set generalizes and outperforms an approach that uses only a subset of those features.Source: A. Context Analysis · p. 5
  • Context integration reduces miss rate by a further 0.5–1% for both ACF and LDCF84 models.The proposed modifications also produce consistent improvements over the ACF and LDCF baselines by about 10%.Source: A. Context Analysis · p. 5
  • On new Caltech annotations, ACF+ and ACF++ are on par with state-of-the-art detectors, while LDCF84++ outperforms the VGG baseline on MRN-4.The table caption describes LDCF84++ as using four filters on top of ACF features and having low computational complexity.Source: Table I · p. 5
  • The authors suggest further gains may come from combining their insights with those from other studies.Source: A. Context Analysis · p. 5

B. Results on Caltech-New

On improved Caltech annotations, the proposed techniques achieve the best detection results reported among CNN and non-CNN methods on the challenging MRN-4 metric.

B. Results on Caltech-New

C. Generalization to Face Detection

The training procedure transfers to face detection on WIDER and FDDB, with performance gains as decision-tree depth increases on WIDER validation and state-of-the-art results reported on FDDB.

C. Generalization to Face Detection

IV. CONCLUDING REMARKS

The experiments report improved face-detection performance on WIDER and FDDB, while concluding that boosted trees have limited representational capacity. Increasing weak-learner complexity saturates early, motivating architectural changes for handling large datasets.

IV. CONCLUDING REMARKS

  • The experiments study the boosted decision-tree model used in many vision tasks.The paper analyzes generalization, modeling capacity, dataset size and balance, and overfitting handling.Source: §IV · pp. 5–6
  • On WIDER, LDCF+ improves markedly over the ACF baseline; on FDDB, results improve over prior HOG+LUV state of the art while competing with CNN approaches.The WIDER comparison covers three difficulty settings; the FDDB comparison uses discrete-score evaluation.Source: Figure 5 · p. 6, Figure 6 · p. 6
  • Increasing weak-learner complexity saturates early, limiting further gains from deeper decision trees even with extensive augmentation.Source: §IV · p. 6
  • Architectural changes at the decision-tree training level are required to increase modeling capacity effectively when handling large datasets.Source: §IV · p. 6
Loading 1701.01692v1…