AI Paper Atlas

Research map / Video / Video-language understanding

Robustness Analysis of Video-Language Models Against Visual and Language Perturbations

Madeline Chantry Schiappa, Shruti Vyas, Hamid Palangi, Yogesh S Rawat, Vibhav Vineet

NeurIPS 2022 · Video › Video-language understanding · approach: Video-language pretraining · 6 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Perturb video rather than textVideo-language retrieval models · MSRVTT-P and YouCook2-Prelative robustness γ^r: MSRVTT-P: 0.91±0.15 vs 0.92±0.13; YouCook2-P: 0.55±0.37 vs 0.91±0.10Hurt
Use pre-training instead of training from scratchVideo-language retrieval models · MSRVTT-P and YouCook2-Probustness and retrieval performanceImproved
Use FIT's ViT-based visual encoder rather than the studied CNN-based visual encodersVideo-language retrieval · MSRVTT-P visual noiserelative robustness γ^r: 0.81±0.21 vs 0.24±0.19Improved
Apply ChangeChar text perturbationsVideo-language retrieval models · MSRVTT-P and YouCook2-Prelative robustness γ^r: MSRVTT-P: 0.71±0.11 to 0.81±0.09; YouCook2-P: 0.75±0.10 to 0.84±0.10Hurt
Perturb video and text together rather than separatelyVideo-language retrieval models · MSRVTT-Prelative robustness γ^r: 1 and 0.98 individually; 0.89 combinedHurt
Fine-tune on the target dataset rather than evaluate zero-shotVideo-language retrieval models · MSRVTT-Prelative robustness γ^r: FIT video: 0.91±0.15 to 0.86±0.15; FIT text: 0.92±0.13 to 0.88±0.10Hurt
Fine-tune on the target dataset rather than evaluate zero-shotVideo-language retrieval models · YouCook2-P video perturbationsrelative robustness γ^r: UniVL: 0.50±0.36 to 0.55±0.36Improved
Use cross-attention rather than a two-branch encoderVideo-language retrieval models · YouCook2-Pabsolute robustnessImproved
Preserve alignment between video segments and their text rather than jumbling segmentsVideo-language retrieval models · YouCook2-ProbustnessHurt
Perturb frame order within a text-aligned segmentVideo-language retrieval models · MSRVTT-P and YouCook2-ProbustnessNo clear effect
Train on the target dataset rather than rely on a larger pre-training datasetVideo-language retrieval models · MSRVTT-P gender-neutral text perturbationsretrieval performanceHurt
Use Word2Vec rather than BERT text featuresZero-shot video-language retrieval · MSRVTT-P DropText, Positional, and ChangeCharrelative robustness γ^r: DropText: 0.57±0.32 vs 0.36±0.26; Positional: 0.78±0.15 vs 0.63±0.15; ChangeChar: 0.77±0.10 vs 0.71±0.10Mixed

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy