Samsung R&D Institute Ukraine's Visual Intelligence Team Wins the Runner-Up Prize at ICCV 2025 Perception Test Challenge

The Perception Test Challenge, held as part of the International Conference on Computer Vision (ICCV) 2025, is a prestigious competition aimed at advancing the state-of-the-art in visual perception tasks. The challenge focuses on several key research problems, advancing multimodal perception in AI systems, particularly in:

Visual Question Answering (VQA): Evaluating AI's ability to answer questions about images accurately.

Object Recognition & Localization: Identifying and precisely locating objects in complex scenes.

Action Recognition: Understanding human actions and interactions in videos.

Scene Understanding: Interpreting contextual relationships between objects and environments.

Robustness to Adversarial Conditions: Ensuring performance under varying lighting, occlusion, and noise.

The Perception Test Challenge attracts a wide range of approaches and methods from top researchers and institutions worldwide. Key open problems include: generalization across domains (e.g., synthetic to real-world data), interpretability of AI decisions in perception tasks, and efficiency in real-time applications (e.g., autonomous systems).

This year, the third Perception Test Challenge received 557 submissions from 81 teams across 5 tracks. Samsung R&D Institute Ukraine has demonstrated top results in Joint Object and Point tracking.

In the joint object & point tracking task, there are two types of queries:

object query: The model receives a video and a bounding box representing an object, and it is required to track the object throughout the video.

point query: The model receives a video and the 2D coordinates of a point, and it is required to track the point throughout the video, also accounting for occlusions.

Our Visual Intelligence Team developed a novel approach that combines advanced CNN architectures with transformer models for enhanced feature extraction and context understanding. In particular, we extended SAM2 segmentation tracker with Point Head for coordinate prediction and applied Fusion Layer in memory to combine mask and point features. We leveraged LocoTrack for fine-grained point tracking and integrated DAM4SAM for temporal consistency. This approach demonstrated high accuracy on small and partially occluded objects, while maintaining point tracking even when segmentation disappeared.

Finally, our method demonstrated superiority in several key metrics, including accuracy and robustness, securing the runner-up prize in the challenge.

Samsung R&D Institute Ukraine's Visual Intelligence Team's achievement in the Perception Test Challenge at ICCV 2025 highlights our commitment to advancing the state-of-the-art in visual perception. The ability to accurately perform both object and point tracking simultaneously is critical for perception systems in robotics, augmented reality, and embodied AI. Precise point tracking and segmentation can enhance AR/VR experiences by enabling more accurate object interaction and spatial understanding. Our innovative approach and dedication to excellence have positioned us as leaders in the field, driving future innovations in real-world applications.

We are proud to be recognized among the winners and look forward to continuing our contributions to the field.

For more details, visit the https://perception-test-challenge.github.io/