PaperlayerView exact revision on arXiv

Source-linked AI summary

MDFN: Multi-Scale Deep Feature Learning Network for Object Detection

Wenchi Ma, Yuanwei Wu, Feng Cen, Guanghui Wang

arXiv:1912.04514v1cs.CV

TL;DR

Object detection needs detail, semantic, and contextual information, yet small and occluded objects remain difficult for deep convolutional networks. MDFN learns multi-scale deep features with efficient inception modules and achieves strong benchmark performance using a relatively small framework. The paper concludes that deep features support accurate localization and classification while maintaining computational efficiency.

  • Problem

    Small and occluded object detection remains insufficiently solved, despite the use of more complicated deep convolutional models.

  • Method

    MDFN learns multi-scale deep features and uses information square and cubic inception modules with parameter sharing in high-level layers.

  • Results

    MDFN outperforms state-of-the-art models on KITTI and achieves more accurate localization and classification results on general objects.

  • Takeaways & Limitations

    Deep features provide semantic and contextual information that supports the detector's performance on difficult objects.

  • Takeaways & Limitations

    On the reported comparison, MDFN-I2 is not higher than MDFN-I1 by mAP, and one discussed dataset represents only a tiny portion of the entire dataset.

Abstract

from arXiv · show

This paper proposes an innovative object detector by leveraging deep features learned in high-level layers. Compared with features produced in earlier layers, the deep features are better at expressing semantic and contextual information. The proposed deep feature learning scheme shifts the focus from concrete features with details to abstract ones with semantic information. It considers not only individual objects and local contexts but also their relationships by building a multi-scale deep feature learning network (MDFN). MDFN efficiently detects the objects by introducing information square and cubic inception modules into the high-level layers, which employs parameter-sharing to enhance the computational efficiency. MDFN provides a multi-scale object detector by integrating multi-box, multi-scale and multi-level technologies. Although MDFN employs a simple framework with a relatively small base network (VGG-16), it achieves better or competitive detection results than those with a macro hierarchical structure that is either very deep or very wide for stronger ability of feature extraction. The proposed technique is evaluated extensively on KITTI, PASCAL VOC, and COCO datasets, which achieves the best results on KITTI and leading performance on PASCAL VOC and COCO. This study reveals that deep features provide prominent semantic information and a variety of contextual contents, which contribute to its superior performance in detecting small or occluded objects. In addition, the MDFN model is computationally efficient, making a good trade-off between the accuracy and speed.

1. Introduction

MDFN addresses object detection by learning semantic and contextual information from deep features while retaining multi-scale representations. Its efficient modules and small base network target improved detection, especially for small and occluded objects.

  • Motivation: Deep networks can attenuate or merge useful features during forward transmission, while detection requires both detail features and semantic context.Semantic context describes object relationships and correlations between objects and their surroundings.
  • Proposed approach: MDFN learns abstract semantic features from concrete features in deep network layers and feeds the extracted information directly to prediction.Multi-scale receptive fields activate contextual information on deep feature maps.
  • Proposed approach: Four inception modules in consecutive deep layers extend feature expression and realize a deep-feature-based multi-scale object detector.The modules extract contextual information while operating within a single-shot framework.
  • Motivation: MDFN focuses on deep features because small and occluded object detection remains difficult for conventional deep convolutional networks.The paper argues that deep features better represent object characteristics and scene semantics.
  • Results: The model uses VGG-16 as a relatively small base network and reports state-of-the-art performance on KITTI with a trade-off between detection accuracy and speed.The paper also reports competitive performance on PASCAL VOC and COCO.
  • Efficiency: Information square and cubic inception modules share filter parameters to reduce model size and computational cost.MDFN also combines outputs from different layer levels to extract multi-scale feature information and shorten transmission paths.

2. Related Work

Related work improves object detection through deeper or wider feature extractors, multi-scale representations, skip connections, and contextual features. These approaches trade feature quality, resolution, accuracy, and computational efficiency in different ways.

  • Feature extraction: Deeper base networks such as ResNet improve feature extraction and representation, but can substantially reduce detection speed.Replacing VGG-16 with Residual-101 improved SSD mAP by 1% while reducing speed from 19 FPS to 6.6 FPS.
  • Feature extraction: VGG-16 remains widely used because its relatively shallow and thin structure offers a trade-off between detection accuracy and running speed.SSD combines VGG-16 with a multi-box detector in an end-to-end network.
  • Network width: Wider architectures such as GoogleNet use inception modules with filters of different sizes to produce concatenated multi-size feature maps.Residual-inception variants also reported advantages over individual inception and residual techniques.
  • Deep features: Deep-feature research identifies redundancy in propagation and reports that most ResNet-101 gradients come from layers 10 to 34.Other methods instead draw multi-scale information from shallow layers because of their detail features.
  • Deep features: Experiments indicate that semantic features and object context contribute to detecting small or occluded objects alongside shallow-layer details.DSSD adds context through deconvolution layers and skip connections, while Mask R-CNN uses fine spatial layouts for mask prediction.
  • SSD: SSD is a single-shot multi-box detector built from a VGG-16 feature extractor and a classification-localization network using default boxes at multiple scales.Its design provides a trade-off between higher detection accuracy and real-time detection efficiency.

3. Deep Feature Learning Network

MDFN extracts and directly uses deep, multi-scale features to combine semantic abstraction, contextual information, and fine details for object detection. Its inception modules and selective feature transmission aim to improve efficiency while supporting localization of difficult objects.

  • MDFN constructs deep feature-learning inception modules in the top four layers to efficiently extract abstract feature representations.
  • The network combines multi-box, multi-scale, and multi-level techniques within a single-shot object-detection framework.
  • Deep feature maps are processed directly and fed to output layers, shortening propagation paths and improving feature-use efficiency.
  • Multi-scale filtering on deep maps increases extracted-information dimensionality and strengthens semantic expression.
  • The scheme combines high-resolution shallow features for fine details with deep features for abstraction and multi-scale receptive fields for varied object contexts.
  • Intermediate-level contextual information supports localization of overlapped, occluded, small, blurred, or saturated objects, while consecutive deep inception improves localization and classification.
  • The design uses information square and cubic inception modules with parameter sharing, and defines MDFN-I1 and MDFN-I2 architectures.

4. Experiments

Across KITTI, PASCAL VOC, and COCO experiments, MDFN delivers strong detection accuracy, particularly for small, occluded, and hard objects, while maintaining an accuracy–efficiency trade-off with VGG-16. Its advantage is especially pronounced at higher IoU thresholds, although MDFN-I2 does not consistently surpass MDFN-I1 in overall mAP.

  • KITTI results: MDFN-I1 ranks second among published KITTI methods, with mAP nearly 4% above SqueezeDet+ and almost 5% above VGG16 + ConvDet.MDFN also performs best for moderate and hard objects across car, pedestrian, and cyclist categories.
  • KITTI results: MDFN improves detection of small and occluded objects, detecting examples that SSD and other models miss in cluttered scenes.Qualitative examples include tiny occluded vehicles, additional pedestrians, and an occluded cyclist.
  • KITTI results: MDFN models perform consistently better than SSD across IoU thresholds, with MDFN-I2 exceeding SSD by 6% for pedestrians at IoU 0.75 and over 8% for cyclists at IoU 0.8.The advantage of MDFN-I2 becomes especially evident above IoU 0.65 for pedestrian and cyclist detection.
  • COCO results: On COCO, MDFN-I2-512 achieves the highest accuracy under AP, AP50, and AP75, while MDFN models outperform both two-stage and one-stage detectors on all three criteria.AP averages precision over 10 IoU levels from 0.5 to 0.95 across 80 categories.
  • Limitations and comparisons: MDFN-I2 does not surpass MDFN-I1 in overall mAP on KITTI and VOC2007, despite its deeper multi-scale feature learning.On PASCAL VOC, the paper attributes the result partly to evaluating only a tiny portion of the entire dataset.
  • Efficiency discussion: Adding deep feature learning modules increases parameters by around 10%, but reduces MDFN running speed by less than 4% and increases FLOPs by around 2%.The paper presents this as a favorable detection accuracy–operational efficiency trade-off relative to deeper or wider alternatives.

5. Conclusion

MDFN is a multi-scale deep feature learning detector that integrates deep-layer inception modules to use semantic and contextual information. It achieves accurate localization and classification across VOC, KITTI, and COCO, including difficult small and occluded objects.

  • MDFN is presented as a novel multi-scale deep feature learning network and the first single-shot detector specifically focused on deep feature learning.The paper states that this approach advances state-of-the-art object detection and classification techniques.
  • MDFN integrates information square and cubic inception modules into deep network layers to exploit semantic and contextual information.These modules support the proposed multi-scale deep feature learning detector.
  • Extensive experiments report accurate localization and classification results on VOC, KITTI, and COCO.The results are described as producing consistent and robust semantic representation.
  • Deep-feature context information plays a vital role in detecting difficult objects such as small and occluded ones.
Loading 1912.04514v1…