acceptodds
Desk rejected at ICLR 2027

Too Many to Judge? MultiDance: A Multi-Task Benchmark for Multi-Target Action Quality Assessment

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Action Quality Assessment (AQA) aims to quantify how well actions are performed, serving as a fine-grained video understanding task with broad applications. However, most studies evaluate a single athlete or team in isolation, overlooking competitions where multiple performers share the stage and interact. Such settings require target-specific assessment that accounts for competitive context and provides interpretable feedback. To address these issues, we introduce Multi-Target AQA (MT-AQA), a task that jointly evaluates and compares multiple competing targets within the same video. We present MultiDance, the first large-scale benchmark for MT-AQA, comprising 1,548 long videos from 99 events with 14,498 annotated performer instances of diverse skill levels and annotations for classification, ranking, scoring, and coaching. We further propose Multi-Target Assessment Coach (MultiAC), a multimodal large language model (MLLM) framework that unifies these tasks through domain knowledge prompts and coarse-to-fine dual-stage adaptation. MultiAC outperforms single-target AQA extensions and general MLLMs on MultiDance, where joint training reveals complementary benefits between single- and multi-target data. Results on EgoExo-Fitness further support coarse-to-fine multi-task adaptation beyond dance.

Primary area:
datasets and benchmarks
Keywords:
Multi-Target understanding, Multi-Task benchmark, Action quality assessment, Group activity understanding

Related papers

Discussion (0)

Sign in to comment.