💻 computer science

Evaluation of Winning Solutions of 2025 Low Power Computer Vision Challenge

This paper outlines the design, evaluation framework utilizing the Qualcomm AI Hub, and winning solutions of the 2025 IEEE Low-Power Computer Vision Challenge across image classification, open-vocabulary segmentation, and monocular depth estimation tracks, while highlighting key trends and future directions for efficient edge vision models.

Original authors: Zihao Ye, Yung Hsiang Lu, Xiao Hu, Shuai Zhang, Taotao Jing, Xin Li, Zhen Yao, Bo Lang, Zhihao Zheng, Seungmin Oh, Hankyul Kang, Seunghun Kang, Jongbin Ryu, Kexin Chen, Yuan Qi, George K Thiruvathukal

Published 2026-04-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zihao Ye, Yung Hsiang Lu, Xiao Hu, Shuai Zhang, Taotao Jing, Xin Li, Zhen Yao, Bo Lang, Zhihao Zheng, Seungmin Oh, Hankyul Kang, Seunghun Kang, Jongbin Ryu, Kexin Chen, Yuan Qi, George K Thiruvathukal, Mooi Choo Chuah

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Great Low-Power Vision Race: How Computers Learned to See on a Budget

Imagine you have a super-smart robot assistant. You want it to recognize a cat, a bus, or a soccer ball instantly. But there's a catch: this robot lives on a tiny battery (like a smartphone or a drone) and has a very small brain (limited memory and processing power). If you give it a giant, hungry brain that needs a power plant to run, the robot will die in seconds.

This is exactly the problem the 2025 Low-Power Computer Vision Challenge (LPCVC) tried to solve. Think of it as a "Formula 1" race, but instead of building the fastest car, teams were building the smartest, most energy-efficient eyes for tiny robots.

Here is a breakdown of the race, the three tracks, and how the winners did it.

The Three Tracks: Three Different Puzzles

The organizers set up three different "obstacle courses" (tracks) to test different skills:

1. The "Chameleon" Challenge (Image Classification)

The Goal: Identify an object (like a plane or a book) even if the lighting is weird, the photo is blurry, or the image is a computer-generated drawing.
The Analogy: Imagine you are playing "I Spy." Usually, you see a red ball in a sunny park. But in this challenge, you have to find that same red ball in a pitch-black cave, under a harsh spotlight, or inside a video game.
The Winner: A team from Ajou University (LabLVM) didn't just memorize what objects look like; they learned the essence of objects. They used a clever trick called "Layer Merging and Imitation."

The Trick: Think of a deep neural network like a multi-story factory. Usually, every floor does a specific job. LabLVM realized they could combine the work of several floors into one super-floor. They taught this new, shorter factory to "pretend" it was the original, taller one. This made the brain much faster without losing its smarts.

2. The "Mind Reader" Challenge (Open-Vocabulary Segmentation)

The Goal: You give the robot a text command like "Find the green ping-pong ball under the other balls," and it must highlight exactly that ball on the screen. The catch? The robot has never seen a "green ping-pong ball" before.
The Analogy: It's like asking a librarian to find a book they've never cataloged before, just by describing the cover. The robot has to understand the meaning of the words and match them to the picture instantly.
The Winner: A team from Southeast University (SICer) won by building a better translator.

The Trick: They realized the robot was spending too much time "thinking" about how to connect words to pictures. They streamlined the process, replacing a heavy, slow thinking engine with a lightweight, fast one. They also trained the robot on a massive library of images and descriptions so it could guess what new things might look like.

3. The "3D Vision" Challenge (Monocular Depth Estimation)

The Goal: Look at a flat 2D photo and guess how far away every object is, creating a 3D map.
The Analogy: Imagine looking at a painting of a forest and trying to tell the painter exactly how many steps away each tree is, without ever walking into the forest.
The Winner: A team from the University of Minnesota (Sailor Moon) won by being a master of "tweaking."

The Trick: They took an existing, powerful 3D vision model and performed "surgery" on it. They found parts of the code that were doing unnecessary math (like doing the same calculation twice) and fused them together. They also found the "Goldilocks" resolution for the images: not too big (too slow) and not too small (too blurry), but just right to get the best score in the shortest time.

The Race Track: How They Measured Success

In most computer competitions, the only thing that matters is: "How many did you get right?"

In this race, the judges used a two-part scorecard:

Accuracy: Did you get the answer right?
Speed & Efficiency: Did you do it fast enough to run on a real phone?

If a team got 100% accuracy but took 10 seconds to think, they were disqualified. It was like a race where you get points for finishing first, but you lose points if you run out of gas.

They used a special cloud platform called Qualcomm AI Hub to test everyone. It was like a universal testing ground where every robot had to run on the exact same "engine" (Snapdragon chips) to ensure it was a fair fight.

The Big Takeaways: What Did We Learn?

The paper concludes with some important lessons for the future of AI:

Don't Reinvent the Wheel: The winners didn't build new brains from scratch. They took existing, famous brains (like MobileNet or Depth-Anything) and optimized them. It's like taking a Ferrari engine and tuning it to run on regular gasoline instead of premium.
Time Matters: The teams that improved the most were the ones who had time to experiment, fail, and try again. The competition lasted long enough for them to iterate.
The Future Needs More: The authors suggest that future competitions should care even more about battery life and memory usage, not just speed. They also want to see more creativity, rewarding teams who invent new ways of thinking, not just teams who tweak existing code.

In a Nutshell

This competition proved that we don't need giant, power-hungry supercomputers to see the world. With smart engineering, we can put "super-vision" into tiny, battery-powered devices that can help us navigate, explore, and interact with the world in real-time. The winners showed us that efficiency is the new intelligence.

Based on the provided paper, here is a detailed technical summary of the 2025 IEEE Low-Power Computer Vision Challenge (LPCVC).

1. Problem Statement

The paper addresses the critical challenge of deploying high-performance computer vision models on resource-constrained edge devices (e.g., mobile phones, embedded systems, drones). While deep learning models have achieved state-of-the-art accuracy, they often fail to meet strict constraints regarding latency, memory capacity, and energy efficiency required for real-time, battery-powered applications.

The 2025 LPCVC aimed to bridge this gap by evaluating solutions that balance accuracy with efficiency across three distinct, emerging vision tasks:

Robust Image Classification: Handling significant variability in lighting (normal, low, backlit) and image styles (indoor, outdoor, synthetic).
Open-Vocabulary Segmentation: Generating segmentation masks based on natural language text prompts for categories not seen during training.
Monocular Depth Estimation: Estimating relative depth from a single RGB image without active sensors or multi-view setups.

2. Methodology and Competition Framework

Competition Infrastructure

Platform: The competition utilized the Qualcomm AI Hub, a cloud-based environment providing access to Snapdragon hardware (Snapdragon 8 Elite and X Elite) for consistent, reproducible benchmarking.
Evaluation Workflow:
1. Participants compiled models via the AI Hub to generate a unique Compile Job ID.
2. IDs were submitted to the official website (lpcv.ai).
3. A referee system at Purdue University coordinated remote inference on actual hardware.
4. Latency was measured to ensure compliance with thresholds, and model outputs were scored against hidden test sets.
Scoring Mechanism: Final scores integrated both accuracy and efficiency. For example, Track 1 used a formula: $Score = \frac{Accuracy}{\max(\frac{Execution Time}{2}, 1)}$ , penalizing excessive latency while rewarding speed.

Track-Specific Methodologies

Track 1: Image Classification
- Dataset: 64 object categories from COCO, annotated with lighting and style labels. Synthetic images were generated via Stable Diffusion.
- Baseline: MobileNetV2 (224×224 input).
- Constraints: Inference latency < 10 ms on Snapdragon 8 Elite.
- Winning Strategy (LabLVM): Utilized MobileCLIP (a Vision-Language Model) but optimized it via Layer Merging and Imitation.
  - Layer Merging: Weighted summation of selected encoder layers to reduce depth while preserving feature magnitude.
  - Layer Imitation: A lightweight training approach where the merged layer mimics the original layer's output using Mean Squared Error (MSE), avoiding full fine-tuning.
Track 2: Open-Vocabulary Segmentation
- Task: Segment objects based on text prompts (e.g., "a green bus") without prior exposure to the specific class.
- Baseline: X-Decoder architecture (Focal-T backbone + CLIP tokenizer).
- Constraints: Inference latency < 1000 ms on Snapdragon X Elite.
- Winning Strategy (SICer):
  - Data Augmentation: Expanded training with Visual Genome and AI-generated prompts; applied contrast/brightness adjustments and occlusion.
  - Architecture: Replaced standard multi-head attention with Multi-Scale Linear Attention (MLA) to reduce FLOPs.
  - Loss Functions: Combined an IoU-based loss (for common objects) with a global mIoU loss (for rare/small objects) and used Online Hard Example Mining (OHEM).
  - Fine-tuning: Two-stage process using LoRA and Visual Prompt Tuning (VPT).
Track 3: Monocular Depth Estimation
- Task: Predict relative depth maps from single RGB images.
- Baseline: Depth-Anything-V2 (Transformer-based).
- Constraints: Inference latency < 34 ms (30 fps) on Snapdragon 8 Elite.
- Winning Strategy (Sailor Moon):
  - Graph Optimization: Identified and fused constant operations (e.g., normalization subtraction/multiplication) into single convolution layers to eliminate runtime overhead.
  - Resolution Tuning: Systematically tested input resolutions to find the optimal trade-off, selecting 434×322 pixels to maximize the F-score within the time limit.

3. Key Contributions

Holistic Evaluation Framework: The paper establishes a rigorous benchmarking protocol that enforces real-world hardware constraints (Snapdragon chips) rather than relying solely on theoretical FLOPs or cloud-based inference.
Novel Optimization Techniques:
- Introduction of Layer Merging and Imitation to compress Vision-Language Models without retraining the entire network.
- Demonstration of Multi-Scale Linear Attention (MLA) for efficient open-vocabulary segmentation.
- Practical application of model graph optimization (fusing constant operations) to meet strict latency budgets.
Comprehensive Analysis of Winning Solutions: The paper provides a deep dive into how top teams (e.g., LabLVM, SICer, Sailor Moon) balanced architectural innovation with resource constraints, moving beyond simple hyperparameter tuning.
Qualcomm AI Hub Integration: The paper details the successful implementation of a cloud-to-edge evaluation pipeline, ensuring that all submissions were tested on identical hardware configurations.

4. Results

The winning solutions demonstrated significant improvements over the provided baselines while adhering to latency constraints:

Track	Metric	Baseline	Top Team	Improvement	Latency (Top Team)
Track 1 (Classification)	Top-1 Accuracy	0.692	0.974 (LabLVM)	+40.7%	1.612 ms (< 10ms limit)
Track 2 (Segmentation)	mIoU	0.461	0.532 (SICer)	+15.4%	515.8 ms (< 1000ms limit)
Track 3 (Depth)	F-score	62.37	83.14 (Sailor Moon)	+33.3%	30.4 ms (< 34ms limit)

Track 1: LabLVM achieved near-perfect accuracy (0.974) by leveraging VLMs and layer merging, significantly outperforming the MobileNetV2 baseline.
Track 2: SICer achieved the highest mIoU (0.532) by effectively handling open-vocabulary tasks through architectural changes (MLA) and advanced data augmentation.
Track 3: Sailor Moon achieved a high F-score (83.14) by optimizing the computational graph and fine-tuning input resolution, proving that small resolution adjustments can yield large performance gains.

5. Significance and Future Directions

Practical Deployment: The competition validates that high-accuracy models can be deployed on edge devices if optimized for specific hardware constraints. The results show that "efficient" does not mean "low performance."
Trends Observed:
- Optimization over Architecture: Winners largely relied on optimizing existing SOTA models (MobileNet, CLIP, Depth-Anything) via pruning, quantization, and graph fusion rather than designing new architectures from scratch.
- Importance of Starter Kits: Comprehensive baseline solutions lowered the entry barrier, allowing teams to focus on innovation.
Recommendations for Future Competitions:
- Diverse Baselines: Provide a spectrum of baseline models (from tiny/energy-efficient to large/accurate) to encourage diverse approaches.
- Expanded Metrics: Move beyond latency to include energy consumption and peak memory usage as primary ranking metrics.
- Hardware Agnosticism: Evaluate on multiple device types to prevent over-optimization for a single chip.
- Innovation Awards: Introduce sub-tracks specifically for novel network designs to reward architectural creativity, not just optimization.

In conclusion, the 2025 LPCVC successfully demonstrated the maturity of the efficient vision community, showcasing that with rigorous hardware-aware optimization, complex vision tasks can be performed reliably on low-power edge devices.