CVPR 2026 Workshop

A Study of Failure Modes in
Two-Stage Human–Object Interaction Detection

Why do HOI models break down — and where does aggregate mAP hide the problem?

* Equal Contribution

Lemeng Wang* 1Qinqian Lei* 2Vidhi Bakshi1Daniel Yi1Yifan Liu1Jiacheng Hou1,
Asher Seng Hao4Zheda Mai1Wei-Lun Chao3Robby T. Tan2Bo Wang5

1The Ohio State University  ·  2National University of Singapore  ·  3Boston University
4Independent Researcher  ·  5University of Mississippi

📄 Paper 💻 Code (coming soon) 📑 BibTeX
Qualitative examples of HOI failure modes

Figure 1. Representative failure modes across four scene configurations — object confusion, interaction confusion, and pairing confusion — all invisible under standard mAP evaluation.


Abstract

Human–object interaction (HOI) detection aims to detect interactions between humans and objects in images. While recent advances have improved performance on existing benchmarks, their evaluations mainly focus on overall prediction accuracy and provide limited insight into the underlying causes of model failures. In particular, modern models often struggle in complex scenes involving multiple people and rare interaction combinations.

In this work, we present a study to better understand the failure modes of two-stage HOI models. Rather than constructing a large-scale benchmark, we decompose HOI detection into multiple interpretable perspectives and analyze model behavior across these dimensions to study different types of failure patterns. We curate a subset of images from HICO-DET organized by human–object–interaction configurations, and analyze model behavior under these configurations. Importantly, high overall benchmark performance does not necessarily reflect robust visual reasoning about human–object relationships.


Dataset Taxonomy

We reorganize the HICO-DET test set into structured subsets that capture different types of ambiguity in human–object interaction configurations. Images are first split by single-person vs. multi-person scenes, then further subdivided along two axes: object relation (same vs. different object instances) and interaction relation (same vs. different interactions), yielding eight categories total.

Single-Person
SPSO — Single Object
SPMO — Multi Object
Multi-Person
A Same Obj · Same Interaction
B Same Obj · Diff Interaction
C Diff Obj (Same Label) · Same
D Diff Obj (Same Label) · Diff
E Diff Obj (Diff Label) · Same
F Diff Obj (Diff Label) · Diff

Categories E & F are rare in HICO-DET (1 and 9 images respectively).


Results

We evaluate four representative two-stage HOI detection models across our structured subsets. All models use official checkpoints without additional fine-tuning.

ADA-CM CMMP HOLa LAIN
Single-person distribution (SPSO vs SPMO)
Figure 3a. SPSO dominates with 5,897 images; SPMO contains only 227. Standard mAP is largely influenced by the simpler SPSO setting.
Multi-person category distribution A-F
Figure 3b. Multi-person category distribution (A–F). Categories A and C are largest; E and F are rare by natural data distribution.
mAP: Single-person vs Multi-person
Figure 4. All four models show a consistent performance drop in multi-person settings, confirming increased scene complexity.
mAP across categories A-D, SPSO, SPMO
Figure 5. Category C consistently shows the lowest mAP across all models — a systematic error tied to instance-level ambiguity.

Key Findings

📉

Multi-Person Gap

All models degrade significantly in multi-person scenes. HICO-DET is >60% single-person, masking this gap in aggregate mAP.

🔀

Category C is Hardest

When multiple objects of the same class are present with different subjects, models fail to correctly pair human–object instances.

🏷️

Verb Errors Dominate

Verb prediction is the most frequent error type. Many high-confidence false positives persist even above 0.5 threshold.

📦

Object-Centric Bias

Models favor verbs that dominate for a given object class — performance correlates with object-conditioned verb frequency, not global HOI frequency.

🔗

Instance Representations

HOLa and LAIN — which use instance-level features rather than region crops — exhibit lower pairing error rates, suggesting better spatial grounding.

📊

Benchmark Limitation

Standard mAP hides category-specific failures. Our structured analysis reveals error sources invisible in aggregated evaluation.


Acknowledgments

We are grateful for the support of the Ohio Supercomputer Center for providing computational resources. This work was supported in part by the Mississippi Impact Grant (MIG), Office for Research and Economic Development, University of Mississippi.

BibTeX

@inproceedings{wang2026diaghoi, title = {A Study of Failure Modes in Two-Stage Human--Object Interaction Detection}, author = {Wang, Lemeng and Lei, Qinqian and Bakshi, Vidhi and Yi, Daniel and Liu, Yifan and Hou, Jiacheng and Mai, Zheda and Seng Hao, Asher and Chao, Wei-Lun and Tan, Robby T. and Wang, Bo}, booktitle = {CVPR Workshop}, year = {2026}, }