Temporally Identity-Aware SSD With Attentional LSTM

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Traditional object detection methods struggle to utilize temporal information in videos.
• Real-time online object detection in videos remains a significant challenge.

Purpose of the Study:

• To propose a novel Temporal Single-Shot Detector (TSSD) for real-world video object detection.
• To enhance detection accuracy and speed by effectively leveraging temporal information.

Main Methods:

• Utilized Convolutional Long Short-Term Memory (ConvLSTM) to integrate pyramidal feature hierarchies across time.
• Developed an attentional ConvLSTM unit with a temporal attention mechanism for background and scale suppression.
• Incorporated an association loss and multistep training for temporal coherence and an online tubelet analysis (OTA) for identification.

Main Results:

• The TSSD-OTA framework demonstrated competitive performance in both detection and tracking on ImageNet VID and 2DMOT15 datasets.
• Achieved fast processing speeds, enabling real-time online object detection.
• Validated effectiveness in a real-world underwater object grasping maneuver.

Conclusions:

• The proposed TSSD-OTA significantly improves video object detection by effectively integrating temporal dynamics.
• Offers a robust and efficient solution for real-time video analysis tasks, including underwater applications.