PaperlayerView exact revision on arXiv

Source-linked AI summary

Pedestrian Detection with Spatially Pooled Features and Structured Ensemble Learning

Sakrapee Paisitkriangkrai, Chunhua Shen, Anton van den Hengel

arXiv:1409.5209v3cs.CVcs.LG

TL;DR

Many detection applications care about performance within a prescribed false-positive range, yet many classifiers do not directly optimize pAUC. This paper combines a structured ensemble that optimizes pAUC with spatially pooled features; experiments demonstrate effectiveness on synthetic data and visual detection tasks.

  • Problem

    Many classifiers do not directly optimize pAUC, although practical detection is often evaluated over a prescribed false-positive range.

    Source: §1 · p. 1

  • Method

    The paper combines a structured ensemble that directly optimizes pAUC over a selected false-positive range with spatially pooled low-level features.

    Source: §1 · p. 1, §1.1 · p. 2

  • Results

    Experiments demonstrate the approach’s effectiveness on synthetic data and visual detection tasks.

    Source: §4 · p. 15

  • Takeaways & Limitations

    Combined with efficient proposal generation, the approach yields a classifier that optimizes average miss rate.

    Source: §4 · p. 14

    The detector achieves a 23.9% average miss rate with a 2-second scan time per 640×480-pixel image, versus soft cascade’s 27.1% with 0.3 seconds per image.

    Source: §3.3 · p. 14

Abstract

from arXiv · show

Many typical applications of object detection operate within a prescribed false-positive range. In this situation the performance of a detector should be assessed on the basis of the area under the ROC curve over that range, rather than over the full curve, as the performance outside the range is irrelevant. This measure is labelled as the partial area under the ROC curve (pAUC). We propose a novel ensemble learning method which achieves a maximal detection rate at a user-defined range of false positive rates by directly optimizing the partial AUC using structured learning. In order to achieve a high object detection performance, we propose a new approach to extract low-level visual features based on spatial pooling. Incorporating spatial pooling improves the translational invariance and thus the robustness of the detection process. Experimental results on both synthetic and real-world data sets demonstrate the effectiveness of our approach, and we show that it is possible to train state-of-the-art pedestrian detectors using the proposed structured ensemble learning method with spatially pooled features. The result is the current best reported performance on the Caltech-USA pedestrian detection dataset.

1 INTRODUCTION

Pedestrian detection is commonly judged in the low-false-positive region, but many classifiers do not directly optimize pAUC. The paper introduces spatially pooled features and a structured ensemble that directly optimizes pAUC over a selected false-positive range.

1 INTRODUCTION

  • For human detection, pAUC over 0.01–1.0 false positives per image summarizes practical performance and is often the primary measure of interest.pAUC measures ROC area between specified false-positive rates.Source: §1 · p. 1

1.1 Main contributions

  • Spatial pooling applied to covariance features and LBP descriptors improves pedestrian-detection accuracy.The paper proposes it as a low-level visual feature approach for pedestrian detection.Source: §1.1 · p. 2
  • The pAUCEnsT ensemble uses structured learning to optimize pAUC over arbitrary false-positive bounds, unlike conventional boosting’s univariate objective.The method uses sample reweighting and a scoring function that orders positive and negative samples for the selected range.Source: §1.1 · p. 2, §1 · p. 2
  • The combined detector is reported to outperform previously reported results and achieve state-of-the-art performance on INRIA, ETH, TUD-Brussels, and Caltech-USA benchmarks.The passage describes combining spatial pooling and structured ensemble learning.Source: §1.1 · p. 2

1.2 Related work

  • Unlike the linear structural SVM in, the proposed method learns an ensemble and directly optimizes pAUC in arbitrary false-positive ranges.The authors describe it as the first principled ensemble method targeting a range not bounded by zero.Source: §1.2 · p. 3

1.3 Notation

  • The scoring function is a weighted sum of binary weak learners, and the objective evaluates pAUC between specified false-positive rates α and β.The notation defines the range with 0 ≤ α < β ≤ 1.Source: §1.3 · p. 4

2 OUR APPROACH

The approach combines spatially pooled low-level features with structured ensemble learning to optimize partial AUC over a user-defined false-positive-rate range. It also refines the feature-map relaxation and uses a two-stage detector to reduce the samples requiring evaluation.

2.1 Spatially pooled features

  • Combining edge orientations O1 and O2 performed slightly better in preliminary experiments than using O1 alone; the mapped input is 9-dimensional, yielding 45 stored covariance values.The covariance descriptor is a symmetric 9 × 9 matrix, so only its upper triangular part is stored.Source: §2.1 · p. 5
  • Spatially pooled LBP (sp-LBP) represents pooled patchwise LBP histograms, using uniform LBP patterns with at most two bitwise transitions to filter noise.The implementation uses a 3 × 3 neighborhood and patch sizes of 4 × 4, 8 × 8, and 16 × 16 pixels.Source: §2.1 · pp. 5–6
  • Max-pooling summarizes covariance descriptors from nearby patches to improve translation and deformation invariance; patchwise covariance computation can use integral images.The method pools vectorized upper-triangular covariance entries and computes correlation coefficients within each patch, enabling parallel feature extraction on a GPU.Source: §2.1 · p. 5
  • Unlike conventional unsupervised feature learning, sp-Cov and sp-LBP pool descriptors directly, avoiding dictionary learning and feature encoding while keeping feature dimensionality lower than visual-word representations.The paper notes that too few visual words can degrade recognition, while too many make features high-dimensional and classifier training computationally infeasible.Source: §2.1 · p. 6

2.2 Optimizing partial AUC

  • The pAUC ensemble optimizes detection performance within the prescribed false-positive-rate interval by learning a weighted sum of binary weak learners.Its structured objective ranks positive-negative pairs in that interval; column generation selects weak learners, and a cutting-plane method optimizes their coefficients.Source: §2.2 · pp. 6–8
  • The pAUC ensemble objective decreases at each iteration as the method adds weak learners and uses cutting planes to find violated constraints.Weak learners are trained as decision trees, and the cutting-plane process continues until no constraint is violated beyond its precision threshold.Source: §2.2 · pp. 7–8
  • Redefining the joint feature map over negative-instance subsets yields a tighter hinge relaxation, faster training, and slightly better pAUC performance than the earlier work.The earlier feature map was computed over all negative instances rather than those ranked within the target false-positive-rate range.Source: §2.2 · pp. 9–10

2.3 Region proposals generation

  • The two-stage detector uses BING to filter background patches before evaluating the proposed features with pAUCEnsT on the remaining samples.The first stage uses fast-to-extract features; the second stage applies the proposed detector to samples that pass the first stage.Source: §2.3 · p. 10
  • The BING pedestrian detector uses 8 × 16 templates, Caltech training data, and multiscale scanning at eight scales per octave.The 8 × 16 template matches the aspect ratio used in, and BING features are stored as UINT128.Source: §2.3 · p. 10
  • Spatial pooling improves the translation invariance of the features, as reported for the comparison with features without pooling.Table 2 reports log-average miss rate for features with and without spatial pooling; no values are provided in the supplied passage.Source: Table 2 · p. 10

3 EXPERIMENTS

Experiments evaluate spatially pooled features, pAUC-optimized ensemble learning, and pedestrian detectors on benchmark datasets. The reported results include strong benchmark performance alongside trade-offs in scanning time.

3.1 Spatially pooled features

  • sp-Cov performs better than M+O+LUV+LBP on INRIA, worse on ETH, and on par on TUD-Brussels; combining sp-Cov, sp-LBP, and M+O+LUV achieves the best performance.The comparison reports log-average miss rates across feature combinations.Source: Table 3 · p. 11, §3.1 · p. 11
  • More than 5% miss-rate reduction at 10^-4 false positives per window was reported for sp-Cov over the original covariance detector.Spatial pooling also reduced miss rate by more than one percent on INRIA and improved robustness to small deformations and translations.Source: §3.1 · pp. 10–11

3.2 Ensemble classifier

  • On protein-interaction prediction, the proposed approach outperforms existing algorithms optimizing either AUC or pAUC.The authors attribute its improvement over SVMtight pAUC [0, 0.1] to introducing nonlinearity, a phenomenon also observed in face detection.Source: Table 4 · p. 11, §3.2 · p. 12
  • On the toy dataset, pAUCEnsT outperforms the other asymmetric classifiers in the FPR range [0, 0.2], while AdaBoost performs worst.The authors note that pAUCEnsT emphasizes positive samples to maximize detection at FPR < 0.2; AdaBoost optimizes overall accuracy.Source: §3.2 · p. 11, Figure 4 · p. 12
  • pAUCEnsT achieves the best performance on all three vision datasets.The comparison reports average pAUC scores in the FPR range [0, 0.1] across USPS digits, scenes, and faces.Source: Table 5 · p. 12, §3.2 · p. 12

3.3 Pedestrian detection

  • On KITTI, the detector outperforms all previously reported benchmark detectors, and on Caltech it outperforms existing detectors by a large margin.Caltech feature selections concentrate on the head, shoulders, and feet; training takes under 24 hours on a parallelized quad-core processor.Source: Figure 5 · p. 13, §3.3 · pp. 13–14, Figure 6 · p. 13
  • At a BING threshold of 0.008, the two-stage detector retains performance similar to pAUCEnsT while halving window-scanning time.Thresholds from −0.004 to 0.008 yield similar detection performance.Source: §3.3 · p. 14, Figure 8 · p. 14
  • The soft cascade performs up to 5% worse in log-average miss rate than the two-stage detector and has higher window-scanning time in the reported comparison.The authors attribute the additional scanning time to repeated comparisons of partial sums against the rejection threshold.Source: §3.3 · p. 14

4 CONCLUSION

The conclusion combines strengthened low-level visual features with a new ensemble-learning method for object detection. Experiments report effectiveness on synthetic data and visual detection tasks, while low-resolution pedestrian detection is identified as future work.

4 CONCLUSION

  • The proposed approach strengthens low-level visual features and formulates a new ensemble-learning method for object detection.Combined with efficient proposal generation, the classifier optimizes average miss rate.Source: §4 · p. 14
  • Experiments demonstrate the approach’s effectiveness on synthetic data and visual detection tasks.Source: §4 · p. 15
  • Applying the approach to a multiple-scales detector is proposed as future work to improve detection of low-resolution pedestrian images.Source: §4 · p. 15

APPENDIX

The appendix analyzes convergence of structured ensemble learning in Algorithm 1. It states that the objective decreases at each iteration and gives a lower bound on the decrease between iterations.

4.1 Convergence analysis of Algorithm 1

  • At each iteration of Algorithm 1, the objective value decreases.Adding a weak learner with a nonzero coefficient reduces the objective; if its coefficient is zero, the solution and objective remain unchanged.Source: §4.1 · p. 16
  • The appendix presents its analysis as establishing convergence of Algorithm 1.Source: §4.1 · p. 16
  • The decrease in objective value between iterations t −1 and t is not less than a specified bound.The appendix introduces this bound as part of its convergence analysis, but the supplied passage does not include its expression.Source: §4.1 · pp. 16–17
Loading 1409.5209v3…