Author ORCID Identifier

 https://orcid.org/0009-0004-5876-5679

Date of Award

Spring 5-20-2026

Document Type

Thesis (Master's)

Department or Program

Computer Science

First Advisor

Nikhil Singh

Second Advisor

John Bell

Third Advisor

Lorie Loeb

Abstract

Modern 2D and 3D human pose estimators silently fail on archival film footage. Their confidence channels do not surface these failures, and no annotated ground truth exists in the domain against which to evaluate or correct them. The result, in pipelines that lift archival 2D detections to 3D for downstream applications, is animation output with jittering limbs, hallucinated skeletons, and anatomically implausible poses. This thesis addresses these failures through three interventions, working from the data foundation upward.

First, we construct a hand-annotated reliability dataset of over 50,000 joint-frame rows across six segments from four films spanning multiple eras (1910, 1960, 2010, 2016), visual styles, and occlusion conditions. Each row records a per-joint reliability label, a reason for any distrust, and an annotator confidence weight.

Second, we train a binary classifier (LightGBM) on this dataset that predicts per-joint reliability at inference time. Across leave-one-film-out cross-validation on three archival films, the classifier achieves weighted don’t-trust precision of 0.84 with recall of 0.74; on Tier 1 anchor joints (hips, shoulders), weighted precision reaches 0.79.

Third, we produce a refined Deep Screens pipeline incorporating reliability-driven corrections: cubic-spline interpolation across flagged joints, Savitzky-Golay temporal smoothing, per-bone canonical length enforcement, and joint angle clamping.

Before-and-after visual comparison of animation quality, with and without the re- liability filter and constraint stack, constitutes the primary evaluative evidence for the contribution.

Download

Share

COinS