InclusiveHuman-10K: Towards Inclusive Human Parsing Beyond the Intact-Limb Assumption

Conference: ECCV 2026
Paper: ECCV Official
Area: Medical Imaging
Keywords: Human Parsing / Disability Inclusion / Residual Limb and Prosthesis Segmentation / Benchmark Dataset / Boundary-Aware Evaluation

TL;DR

InclusiveHuman-10K (IH-10K) addresses the long-standing "intact-limb assumption" in human body parsing by establishing the first inclusive benchmark featuring 10,555 images, 14,939 person instances, and 160,133 pixel-wise masks across residual limbs, three tiers of prostheses, and mobility assistive devices alongside a subgroup-resolved evaluation protocol.

Background & Motivation

Human body parsing serves as a foundational building block for human-centric vision systems, providing pixel-level semantic part labels that underpin crucial downstream applications including virtual try-on, augmented reality interaction, animation, and clinical rehabilitation and gait analysis. However, prominent benchmarks such as LIP, CIHP, and MHP v2 implicitly assume intact anatomies, defining part categories exclusively for hair, face, symmetric natural limbs, and clothing. According to World Health Organization guidance, approximately 0.5% of the global population requires prosthetic or orthotic care at any given time—representing tens of millions of individuals worldwide. In the absence of explicit category definitions for non-intact anatomies, current state-of-the-art models systematically absorb residual limbs and prosthetic components into the "background" or misassign them to adjacent intact-limb classes during training and testing, producing a severe representational and fairness gap.

This limitation is not merely an ethical oversight; it directly degrades the reliability of assistive, rehabilitation, and clinical workflows that require robust part-level anatomical parsing. Non-intact human bodies break canonical skeletal symmetry and standard topological continuity: limbs may terminate at arbitrary anatomical levels, interface with sockets, or connect to running blades, passive hooks, or mechanical hinges. Because existing models are trained exclusively on intact body priors, they fail to discern residual-limb boundaries or handle human-device interfaces, causing catastrophic failures in real-world human-centric vision pipelines.

This paper bridges this critical blind spot by challenging the default assumption of physical completeness. Core idea: dismantle the intact-limb assumption by establishing InclusiveHuman-10K, the first large-scale human parsing benchmark centered on individuals with limb deficiencies, introducing explicit residual-limb and prosthesis taxonomies alongside subgroup-resolved evaluation across common and disability-related classes.

Method

Overall Architecture

InclusiveHuman-10K establishes an end-to-end data pipeline and benchmarking methodology tailored for non-intact body anatomies. The methodology comprises four primary phases: diverse multi-domain image sourcing across everyday, sports, and clinical contexts; an inclusive three-tier category taxonomy defining residual limbs, prostheses, and mobility aids; a rigorous multi-pass annotation and quality-assurance protocol; and a subgroup-resolved evaluation protocol decoupling common and disability-related metrics with boundary-sensitive measurement.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}, 'subGraphTitleMargin': {'top': 8, 'bottom': 16}}}%%
flowchart TD
    A["Multi-Domain Image Sourcing & Curation<br/>Everyday, sports, and clinical scenes"] --> B["Three-Tier Inclusive Category Schema<br/>27 common + 34 disability-related classes"]
    B --> C["Multi-Pass Verification Annotation Protocol<br/>X-AnyLabeling polygons + blind review + lead arbitration"]
    C --> D["Subgroup-Resolved & Boundary-Aware Evaluation Protocol<br/>All / Common / Disability-Related splits"]
    D --> E["Benchmarking & Bottleneck Analysis<br/>Fine-tuned parsing, semantic segmentation, zero-shot models"]

Key Designs

1. Three-Tier Inclusive Category Schema: eliminating structural omission in parsing taxonomies

To overcome the systemic exclusion where disability-related anatomy is merged into background or confused with intact limbs, IH-10K preserves 27 common categories from standard taxonomies (such as face, hair, intact upper arm, intact calf, pants, upper clothes) while extending the label space with 34 explicit disability-related classes. This schema operates across three core dimensions: first, residual-limb granularity, labeling upper arm, forearm, thigh, and calf on both left and right sides (8 distinct segments) to establish unambiguous anatomical boundaries; second, structural prosthesis taxonomy, categorizing prosthetic devices into cosmetic prostheses (passive appearance restoration without active joints), functional non-articulated prostheses (single-function rigid components such as carbon-fiber running blades or hooks), and articulated prostheses (devices featuring mechanical knee, ankle, or elbow hinges); third, mobility assistive devices, explicitly annotating wheelchairs and crutches. This taxonomy forces models to learn non-canonical anatomical layouts and device-body contact interfaces.

2. Multi-Pass Verification Annotation Protocol: ensuring pixel-accurate boundaries on complex interfaces

Given that residual-limb termination contours, prosthetic sockets, and mechanical pylons exhibit high morphological variability and subtle transitions against clothing and skin, standard automated pseudo-labeling is inadequate. IH-10K recruited 25 annotators trained on clinical anatomy and prosthesis types, who generated precise polygon masks using X-AnyLabeling. When boundaries were partially obscured by clothing or devices, annotators completed contours along the shortest anatomically plausible path; when severe occlusion prevented definitive prosthesis subtyping, annotators followed a disambiguation rule to omit fine-grained type assignments rather than introduce label noise. The quality-assurance pipeline enforces a three-stage review: every mask undergoes full-pass blind secondary review with flagged revisions, followed by targeted correction by the original annotator and final cross-batch spot checks and arbitration by a lead reviewer, with complete audit trail logging.

3. Subgroup-Resolved and Boundary-Aware Evaluation Protocol: unmasking performance disparities on underrepresented classes

Because aggregate performance metrics are overwhelmingly dominated by frequent common classes and obscure severe failures on minority categories, IH-10K introduces a subgroup-resolved evaluation framework. Results are reported separately across All, Common (27 classes), and Disability-Related (34 classes). Beyond standard mean Intersection-over-Union (mIoU), part-based Average Precision (APp50, APpvol), and Percentage of Correctly Parsed semantic Parts (PCP), the protocol incorporates Boundary Intersection-over-Union (BIoU):

\[\text{BIoU} = \frac{|(G_d \cap G) \cap (P_d \cap P)|}{|(G_d \cap G) \cup (P_d \cap P)|}\]

where \(G\) and \(P\) denote ground-truth and predicted masks, and \(G_d\) and \(P_d\) denote dilated boundary bands controlled by width parameter \(d\). BIoU heavily penalizes spatial boundary deviations along the narrow interface of stump-socket and prosthesis-limb transitions, quantifying geometric localization errors that standard area overlap tends to underestimate.

Loss & Training

Supervised models on IH-10K are trained using a standardized 7:1:2 image-level random partition (7,388 train, 1,055 validation, 2,112 test). Specialized human parsing architectures (AIParsing, UniParser) optimize joint cross-entropy and part-correlation or edge losses. General semantic segmentation models (FCN, SegFormer, Mask2Former, SegNeXt) are trained using multi-class cross-entropy combined with Dice or mask losses. Zero-shot open-vocabulary models (OVIS, OVSeg, PromptSeg) are evaluated with frozen pre-trained weights to assess out-of-the-box transfer and prompt grounding on non-intact concepts.

Key Experimental Results

Main Results

The table below reports fine-tuned body parsing and supervised semantic segmentation performance on the official IH-10K test split, with breakdown scores formatted as All (Common / Disability-Related).

Model Architecture	Backbone	mIoU (All | Common / Disab.)	APp50 (All | Common / Disab.)	BIoU (All | Common / Disab.)	PCP (All | Common / Disab.)
AIParsing	ResNet50	30.97 (46.34 / 17.51)	31.70 (46.51 / 6.93)	-	35.92 (48.06 / 18.81)
AIParsing	ResNet101	31.49 (50.78 / 15.60)	46.77 (59.73 / 13.37)	-	48.81 (58.85 / 29.62)
UniParser	ResNet50	39.06 (56.55 / 25.18)	43.08 (65.07 / 25.62)	-	44.66 (65.16 / 28.38)
UniParser	ResNet101	41.43 (58.10 / 28.19)	45.67 (67.77 / 28.11)	-	47.11 (66.76 / 31.50)
FCN	ResNet101	21.14 (35.32 / 9.47)	-	12.82 (19.53 / 7.29)	-
SegFormer	MIT-B5	24.54 (39.73 / 12.04)	-	14.94 (22.23 / 8.94)	-
Mask2Former	Swin-L	28.07 (45.05 / 14.09)	-	19.41 (28.67 / 11.78)	-
SegNeXt	Mscan-l	28.19 (45.01 / 14.34)	-	17.78 (26.12 / 10.92)	-

Ablation Study

To diagnose the capabilities and bottlenecks of zero-shot vision-language models on disability-related semantics, the table below compares open-vocabulary instance segmentation (OVIS), open-vocabulary semantic segmentation (OVSeg), and promptable segmentation (PromptSeg) across varying visual and textual prompt inputs.

Task Setting	Method	Backbone / Prompt	All mIoU	Common mIoU	Disab. mIoU	Diagnostic Note
OVIS (Text Prompt)	Grounded-SAM	Huge / Text	6.10	11.56	1.77	Text grounding collapses on disability concepts
OVIS (Text Prompt)	Grounded-SAM2	Large / Text	4.69	8.97	1.30	Missed text grounding accounts for 85.26% of failures
OVIS (Text Prompt)	X-Decoder	Focal-L / Text	1.04	1.37	0.79	Extreme semantic misalignment on prostheses
OVSeg (Text Prompt)	CLIPSeg	- / Text	2.68	5.81	0.19	Severe catastrophic drop on disability classes
PromptSeg (Point Prompt)	SAM2	Large / Centroid Point	23.63	19.66	26.78	Point prompts lead to severe leakage into adjacent skin/clothing
PromptSeg (Oracle Box)	SAM2	Large / Ground-Truth Box	75.20	70.17	79.36	Mask decoder traces boundaries well once localized
Cascaded System	Grounded-SAM2 + SAM2	Large / Predicted Boxes	4.94	8.85	1.48	Upstream visual-language grounding is the primary bottleneck

Key Findings

• Severe disparity between common and disability classes: Across all fine-tuned architectures, performance on disability-related classes is roughly one-third to two-fifths of that on common classes (e.g., UniParser ResNet101 achieves 28.19 vs 58.10 mIoU; APp50 drops sharply from 67.77 to 28.11). Even SOTA segmentation models such as Mask2Former-Swin-L stall at 14.09 mIoU on disability categories.
• Amplified boundary degradation: When evaluated with BIoU, scores across all models plunge by 35% to 50% relative to standard mIoU (e.g., Mask2Former reaches only 11.78 BIoU on disability classes), demonstrating pronounced geometric shifting and edge leakage around stump-socket transitions.
• Backbone scaling fails to resolve topological deficits: Scaling backbones (ResNet50 to ResNet101, or Swin-T to Swin-L) yields substantial gains on common classes, but benefits on disability classes remain marginal (Swin-T to Swin-L improves common mIoU by 1.93 but disability mIoU by only 1.83), confirming that existing models lack inductive priors for non-intact topologies.
• Open-vocabulary models suffer severe conceptual blind spots: Pre-trained vision-language models fail almost completely when prompted with disability terms (Disability AP below 0.55; 85.26% text grounding failure in Grounded-SAM2). In contrast, SAM2 achieves 79.36 mIoU under ground-truth box prompts, confirming that low-level pixel boundary tracing is viable and the fundamental failure lies in language-driven visual concept grounding.

Highlights & Insights

• First-ever dismantling of the intact-limb assumption: Redefines the human parsing benchmark landscape by establishing residual limbs, multi-tier prostheses, and mobility devices as first-class semantic labels across 10.5K in-the-wild images.
• Dissecting foundation model biases: Reveals a stark dichotomy in modern foundation models: while low-level segmentation decoders (SAM2) can successfully delineate stump and device contours given spatial prompts, high-level vision-language backbones completely fail to ground disability concepts, pinning down the exact bottleneck for future research.
• Subgroup-resolved and boundary-aware evaluation protocol: The decoupled All/Common/Disability reporting structure combined with Boundary IoU prevents long-tail failure modes from being obscured by majority-class averages, providing a template for inclusive computer vision benchmarks.

Limitations & Future Work

• Limitations acknowledged by authors: Mobility assistive devices in IH-10K are primarily limited to wheelchairs and crutches, omitting more complex modern devices such as robotic exoskeletons. Furthermore, severe occlusion occasionally complicates manual fine-grained prosthesis subtyping.
• Identified limitations: The image corpus is sourced from public web datasets oriented toward whole-body poses and sports/daily activities, with relatively fewer close-up macro views of clinical surgical stumps or socket skin ulcers, leaving a domain gap for direct clinical diagnosis.
• Future research directions: Developing parsing architectures with dynamic skeletal topology adaptation, formulating boundary-focused loss functions tailored for stump-device interfaces, and pre-training multi-modal models with medically informed text corpora for improved open-vocabulary disability grounding.

Related Work & Insights

• vs Conventional Human Parsing Benchmarks (LIP / CIHP / MHP v2): Conventional datasets are large but strictly assume intact human anatomy, treating prostheses and stumps as noise or background; IH-10K provides the first fine-grained, inclusive benchmark addressing this gap.
• vs Disability Pose Benchmarks (LDPose / InclusiveVidPose): Pose estimation benchmarks focus on sparse keypoints or coarse limb tracking; IH-10K provides dense, pixel-accurate polygon masks indispensable for virtual try-on, apparel fit, and device interaction analysis.
• vs Open-Vocabulary Segmentation Benchmarks (COCO-Stuff / ADE20K): Standard vision-language pre-training datasets lack fine-grained assistive device vocabulary; IH-10K offers a rigorous benchmark to audit the fairness, inclusion, and generalization of foundation models.

Rating

• Novelty: ⭐⭐⭐⭐⭐ Establishes the first human parsing benchmark breaking the intact-limb assumption with explicit residual-limb and prosthesis taxonomies.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ Thoroughly evaluates fine-tuned parsers, standard semantic segmenters, and zero-shot open-vocabulary models across 10+ baselines with subgroup-resolved metrics.
• Writing Quality: ⭐⭐⭐⭐⭐ Exceptionally articulate motivation, rigorous ethical framing, clear taxonomy definitions, and comprehensive empirical analyses.
• Value: ⭐⭐⭐⭐⭐ Highly impactful for advancing fairness and inclusion in human-centric vision, with immediate utility for assistive robotics, rehabilitation, and virtual avatar generation.