Research map / Video / Action recognition

Benchmarking the Robustness of Temporal Action Detection Models Against Temporal Corruptions

CVPR 2024 · Video › Action recognition · approach: Fully supervised temporal action detection · 11 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Builds THUMOS14-C and ActivityNet-v1.3-C temporal-corruption benchmarks, benchmarks seven TAD models, and proposes FrameDrop augmentation plus a Temporal-Robust Consistency loss to improve robustness and clean mAP.

Problem

Temporal action detection models are not robust to temporal corruptions such as missing, blacked-out, or blurred frames within action instances, and no benchmark existed to measure this.

Approach

Corrupt a small subset of frames (1, 5, 10% of central action frames) with five real-world corruption types to break temporal continuity and create benchmarks for TAD; defend with FrameDrop augmentation (dropping frames in action and background segments) and a TRC loss that aligns tIoU distributions of predictions on clean and corrupted videos via action-centric sampling.

Key results

• Relative robustness of TAD models on THUMOS14-C ranges from 54.16% (E2E-TAD) to 84.31% (TriDet with I3D), with mAP drops of 9.62% to 25.86%.
• TriDet with VideoMAEv2 improves from 61.10 to 68.28 corrupted mAP and relative robustness from 81.29% to 90.31% with FrameDrop and TRC loss, while clean mAP rises from 75.16 to 75.60.
• Localization error grows from 5.5% to 8.1% on corrupted data while wrong-label error stays roughly constant (0.6% vs 0.5%).

Measured effects reported in the paper

Surprising findings

• Corrupting frames near action boundaries sometimes improved performance above clean levels, which the authors attribute to the models treating the black frame as a boundary cue.
• The proposed training strategy improves clean-data mAP in most cases despite being designed for robustness.

Limitations

• Corruptions are limited to a fixed set of five types and three severity levels, with corruption length fixed at one frame in the training method.
• Robustness evaluation is restricted to THUMOS14 and ActivityNet-v1.3 and to seven TAD models.
• Benchmark evaluation of corrupted frames uses a subset of 213 THUMOS14 test videos.

Experimental setup

Datasets: THUMOS14 ActivityNet-v1.3

Benchmarks: THUMOS14-C ActivityNet-v1.3-C

Metrics: mAP relative robustness

Baselines: ActionFormer TriDet TemporalMaxer AFSD BasicTAD E2E-TAD VSGN

Builds on: ActionFormer TriDet DETAD Kullback-Leibler divergence

Closest papers by meaning

• A Large-Scale Robustness Analysis of Video Action Recognition Models (CVPR 2023 · similarity 0.89)
• End-to-End Temporal Action Detection with 1B Parameters Across 1000 Frames (CVPR 2024 · similarity 0.88)
• TE-TAD: Towards Full End-to-End Temporal Action Detection via Time-Aligned Coordinate Expression (CVPR 2024 · similarity 0.87)
• Dual DETRs for Multi-Label Temporal Action Detection (CVPR 2024 · similarity 0.87)
• HyrCap: Hybrid Rank-Calibration of Action Proposals for Temporal Event Understanding (NeurIPS 2026 · similarity 0.86)
• Energy vs. Noise: Towards Robust Temporal Action Localization in Open-World (AAAI 2025 · similarity 0.86)