Projects & Publications | TensorsOfTheWall

Person Re-Identification by Analyzing Dynamic Variations in Gait Sequences

ETCCS 2020

Overview

Gait recognition identifies individuals by analyzing their walking patterns but often struggles with appearance changes. Traditional methods like Gait Energy Image (GEI) lose dynamic motion details.

We propose a novel approach using Active Energy Image (AEI), which better captures motion variations. AEIs are divided into segments, and affine moment invariants are extracted as features. Matching is performed using a weighted similarity scheme across all segments.

Evaluated on the CASIA-B dataset, our method achieves 91.13% accuracy with 23 segments and 5 moment invariants, showing strong performance in appearance-variant or low-resolution settings.

Unified Local-Cloud Decision-Making via Residual Reinforcement Learning

K. Sengupta, Z. Shangguan, S. Bharadwaj, S. Arora, E. Ohn-Bar, R. Mancuso

ECCV 2024

Overview

Embodied vision systems like mobile robots must balance energy, latency, and safety. While offloading to the cloud allows access to powerful models, it introduces communication delays that hinder real-time, safety-critical use.

We propose UniLCD, a hybrid inference framework enabling dynamic local-cloud collaboration via a reinforcement learning–optimized routing module and multi-task objective. Tested on a challenging navigation task, UniLCD outperforms state-of-the-art split computing and early exit baselines by over 35%, offering a robust solution for efficient and safe decision-making in dynamic environments.

Person Re-Identification by Analyzing Dynamic Variations in Gait Sequences

S. Bharadwaj, K. Chanda

ETCCS 2020

Overview

Gait recognition identifies individuals by analyzing their walking patterns but often struggles with appearance changes. Traditional methods like Gait Energy Image (GEI) lose dynamic motion details.

We propose a novel approach using Active Energy Image (AEI), which better captures motion variations. AEIs are divided into segments, and affine moment invariants are extracted as features. Matching is performed using a weighted similarity scheme across all segments.

Evaluated on the CASIA-B dataset, our method achieves 91.13% accuracy with 23 segments and 5 moment invariants, showing strong performance in appearance-variant or low-resolution settings.

Unified Local-Cloud Decision-Making via Residual Reinforcement Learning

K. Sengupta, Z. Shangguan, S. Bharadwaj, S. Arora, E. Ohn-Bar, R. Mancuso

ECCV 2024

Overview

Embodied vision systems like mobile robots must balance energy, latency, and safety. While offloading to the cloud allows access to powerful models, it introduces communication delays that hinder real-time, safety-critical use.

We propose UniLCD, a hybrid inference framework enabling dynamic local-cloud collaboration via a reinforcement learning–optimized routing module and multi-task objective. Tested on a challenging navigation task, UniLCD outperforms state-of-the-art split computing and early exit baselines by over 35%, offering a robust solution for efficient and safe decision-making in dynamic environments.

Person Re-Identification by Analyzing Dynamic Variations in Gait Sequences

S. Bharadwaj, K. Chanda

ETCCS 2020

Overview

Gait recognition identifies individuals by analyzing their walking patterns but often struggles with appearance changes. Traditional methods like Gait Energy Image (GEI) lose dynamic motion details.

We propose a novel approach using Active Energy Image (AEI), which better captures motion variations. AEIs are divided into segments, and affine moment invariants are extracted as features. Matching is performed using a weighted similarity scheme across all segments.

Evaluated on the CASIA-B dataset, our method achieves 91.13% accuracy with 23 segments and 5 moment invariants, showing strong performance in appearance-variant or low-resolution settings.

Click on the title in each card to view the full details (link opens in new tab).

Overview