Weakly Aligned Cross-Modal Learning for Multispectral Pedestrian Detection

Abstract

Multispectral pedestrian detection has shown great advantages under poor illumination conditions, since the thermal modality provides complementary information for the color image. However, real multispectral data suffers from the position shift problem, i.e. the color-thermal image pairs are not strictly aligned, making one object has different positions in different modalities. In deep learning based methods, this problem makes it difficult to fuse the feature maps from both modalities and puzzles the CNN training. In this paper, we propose a novel Aligned Region CNN (AR-CNN) to handle the weakly aligned multispectral data in an end-to-end way. Firstly, we design a Region Feature Alignment (RFA) module to capture the position shift and adaptively align the region features of the two modalities. Secondly, we present a new multimodal fusion method, which performs feature re-weighting to select more reliable features and suppress the useless ones. Besides, we propose a novel RoI jitter strategy to improve the robustness to unexpected shift patterns of different devices and system settings. Finally, since our method depends on a new kind of labelling: bounding boxes that match each modality, we manually relabel the KAIST dataset by locating bounding boxes in both modalities and building their relationships, providing a new KAIST-Paired Annotation. Extensive experimental validations on existing datasets are performed, demonstrating the effectiveness and robustness of the proposed method. Code and data are available at: https://github.com/luzhang16/AR-CNN .

Attention Fusion for One-Stage Multispectral Pedestrian Detection

“…We evaluated the proposed fusion method on the KAIST testing dataset in the reasonable setting and in all settings in comparison with ACF+T+THOG [ 14 ], Halfway Fusion [ 22 ], Fusion RPN+BDT [ 30 ], IAF R-CNN [ 28 ], IATDNN+IASS [ 27 ], CIAN [ 37 ], MSDS-RCNN [ 17 ], ARCNN [ 38 ], MBNet [ 39 ], and FusionCSPNet [ 18 ]. Among these detection methods, FusionCSPNet and our method were one-stage methods, and the rest were two-stage methods.…”

Deep Active Learning from Multispectral Data Through Cross-Modality Prediction Inconsistency

“…Miss Rate (lower, better) All Day Night ACF [1] 47.32% 42.57% 56.17% Halfway Fusion [21] 25.75% 24.88% 26.59% Fusion RPN+BF [5] 18.29% 19.57% 16.27% IAF R-CNN [10] 15.73% 14.55% 18.26% IATDNN+IASS [9] 14.95% 14.67% 15.72% CIAN [7] 14.12% 14.77% 11.13% MSDS-RCNN [6] 11.34% 10.53% 12.94% AR-CNN [18] 9.34% 9.94% 8.38% MBNet [12] 8.13% 8.28% 7.86% Ours (full dataset) 8.86% 10.01% 6.77% Ours (10.26% of data) 9.32% 10.13% 7.70% [3] 39.7% 36.1% 36.8% FuseNet [14] 45.6% 41.0% 43.9% RTFNet [13] 53.2% 45.8% 54.8% Ours (full dataset) 53.6% 46.8% 53.3% Ours (17.99% of data) 51.0% 46.6% 48.9% Table 3. mIoU comparisons on TOKYO Dataset.…”

“…This well-known multispectral dataset is built for the pedestrian detection task. In order to tackle the misalignment problem between visible-thermal image pairs, [18] proposes the "paired" annotations by separately relabelling pedestrians for each modality. We remove unpaired images according to the matching of visible and thermal annotations, thus keeping 11,695 images for training.…”

Attention Based Multi-Layer Fusion of Multispectral Images for Pedestrian Detection

“…An accumulated probability fusion layer was also introduced to combine probabilities from different modalities at the proposal-level. Taking the position shift problem of multispectral data into consideration, Zhang et al [54] proposed a region feature alignment module to capture position shifts and a confidence-aware fusion method to merge both modalities.…”