Deep People Detection: A Comparative Study of SSD and LSTM-decoder

Md Atiqur Rahman, Prince Kapoor, Robert Laganière
University of Ottawa
Ottawa, ON, Canada		 Daniel Laroche, Changyun Zhu, Xiaoyin Xu, Ali Ors
NXP Semiconductors
Ottawa, ON, Canada
Questions? Drop us a line

Overview

This study seeks to provide an extensive comparison between two state-of-the-art deep object detection frameworks, namely SSD [1] and LSTM-decoder [2] in the context of people detection. The aim is to explore the two detection architectures in terms of accuracy, speed, robustness to occlusion and scale, as well as generalization ability which are the leading challenges for people detection. Our experimental results show that while the LSTM-decoder can be more accurate in realizing smaller head instances, especially in the presence of occlusions, the sheer detection speed and superior generalization ability of SSD makes it an ideal choice for real-time people detection.

This work was accepted in the 15th Conference on Computer and Robot Vision (CRV'18) conference.


Click here for more

 

Model Architectures

The following gives an architectural overview of LSTM-decoder and SSD models that we study in this work.

Model Pipeline for LSTM-decoder and SSD

Datasets

In addition to using a benchmark head detection dataset called Brainwash [2], we also experimented with two new datasets called MrSub and Clifton which we collected from different restaurants. These datasets vary widely from each other in terms of scale and occlusion level as depicted in the below figure, thereby allowing us to better explore the models' robustness to variation in scale and occlusion.

Datasets

Experimental Setup

The models are evaluated and compared based on 4 different experimental settings as below:

  • Case 1: Models trained and tested on MrSub to analyze the baseline performance. MrSub, as obvisous from the above figure, is the easiest among the three datasests with large scale and very less occlusion.

  • Case 2: Models trained and tested on Brainwash to analyze the robustness to occlusion and tiny-scale. Brainwash is a highly occluded and tiny-scale dataset.

  • Case 3: Models trained on Brainwash, tested on MrSub to analyze the generalization ability over different scales. Brainwash and MrSub clearly have different distributions of scales.

  • Case 4: Models trained on Brainwash+MrSub, tested on Clifton to analyze the domain adaptation ability. Clifton has a distribution of scale and occlusion level which is a mixture of Brainwash and MrSub.

Results

The Recall vs. (1-Precision) curves corresponding to the 4 different cases are depicted below. For each model architecture, we plot 3 different curves corresponding to three different CNN feature extractors (e.g., Inception-Resnet-V2, Inception-V2, and Mobilenet)

Recall vs. (1-Precision) Curves for Case 1 and Case 2

Recall vs. (1-Precision) Curves for Case 3 and Case 4


Some of the sample detections along with the inference time for the two models are shown below.

Sample Detections and Inference Time

References

  1. W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. E. Reed, C.-Y. Fu, and A. C. Berg, “Ssd: Single shot multibox detector,” in ECCV, 2016.
  2. R. Stewart and M. Andriluka, “End-to-end people detection in crowded scenes,” CoRR, vol. abs/1506.04878, 2015.

Navigation

Industry Partners

Thank you to NXP Semiconductor for making this work possible!

NXP