Instruct-ReID++: Towards Universal Purpose Instruction-Guided Person Re-Identification

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Person re-identification (ReID) research has focused on specific tasks, limiting real-world applicability.
• Existing ReID models struggle with diverse scenarios like clothes-changing or visible-infrared matching.

Purpose of the Study:

• To introduce a general ReID task, instruct-ReID, unifying multiple existing ReID tasks into a single model.
• To develop a versatile ReID model capable of retrieving images based on textual or visual instructions.

Main Methods:

• Proposed the instruct-ReID task, treating 6 existing ReID tasks as special cases via instructions.
• Introduced the OmniReID++ benchmark for large-scale, diverse ReID data and evaluation.
• Developed IRM (Instruction-based Retrieval Model) with adaptive triplet loss and IRM++ with memory bank-assisted learning.

Main Results:

• IRM and IRM++ demonstrated superior performance across 10 test sets on the OmniReID++ benchmark.
• Achieved state-of-the-art results in both task-specific and task-free evaluation settings.
• The proposed methods effectively handle diverse retrieval tasks within a unified framework.

Conclusions:

• Instruct-ReID offers a generalized approach to person re-identification, enhancing model applicability.
• The proposed IRM and IRM++ models represent significant advancements in unified ReID.
• The OmniReID++ benchmark facilitates future research in generalized ReID tasks.