Case Study · AI / ML
Human Pose Estimation (HRNet)
Undergraduate thesis on high-precision human pose estimation — keeping high-resolution representations alive through the whole network instead of recovering them from a downsampled bottleneck.
The problem
Human pose estimation is the task of locating body keypoints — shoulders, elbows, knees and so on — in an image. The accuracy of those keypoints is everything for downstream uses like motion analysis, AR, and activity recognition.
The classic encoder–decoder approach downsamples an image to a low-resolution feature map and then tries to upsample back to precise locations. That recover-from-low-res step is exactly where spatial precision leaks away, which caps how accurately fine keypoints can be placed.
How it works
High-resolution representation learning
HRNet's core idea is to never throw the high-resolution signal away. Instead of a single high→low→high path, it runs parallel branches at multiple resolutions from start to finish, keeping a high-resolution stream alive throughout the network.
Repeated multi-scale fusion
The parallel branches repeatedly exchange information, so the high-resolution branch is continually enriched by the semantic context the lower-resolution branches capture. This multi-scale fusion is what lets the network be both spatially precise and semantically aware.
I trained and evaluated on the COCO and MPII keypoint benchmarks, the standard datasets for this task, to validate keypoint precision.
Key decisions & tradeoffs
Maintain parallel high-resolution branches end-to-end
Why: Preserving spatial detail throughout the network avoids the precision loss inherent in downsample-then-upsample architectures.
Tradeoff: Keeping high-resolution branches alive is more compute- and memory-intensive than a single encoder–decoder path.
Validate on both COCO and MPII
Why: Using two standard benchmarks makes results comparable to published work and tests generalization across datasets rather than overfitting to one.
Tradeoff: Two evaluation pipelines and metric conventions (OKS/AP vs PCK) to set up and reason about.
Outcome
- Demonstrated improved keypoint precision by retaining high-resolution representations rather than reconstructing them.
- Worked end-to-end with the standard COCO and MPII benchmarks and their evaluation metrics.
- Built research fundamentals: reading the literature, reproducing a state-of-the-art architecture, and reasoning about evaluation rigor.
What I'd improve next
- Explore lighter-weight variants to trade a little accuracy for real-time inference.
- Extend from single-person to robust multi-person estimation in cluttered scenes.