AI Paper Atlas

Research map / Video / Video-language understanding

CLIP-It! Language-Guided Video Summarization

Medhini Narasimhan, Anna Rohrbach, Trevor Darrell

NeurIPS 2021 · Video › Video-language understanding · approach: Dense video captioning · 15 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
condition frame scoring on dense video captions using cross-modal attention (ablation)CLIP-It · SumMe and TVSum; standard, augment, and transfer configurationsF1: CLIP-Image+Transformer -> CLIP-Image+Video Caption+Transformer: SumMe 53.5 -> 54.2, 55.3 -> 56.4, 51.0 -> 51.9; TVSum 65Improved
use a Transformer instead of a bidirectional LSTM for frame scoring (ablation)CLIP-It · SumMe and TVSum generic summarizationF1Improved
replace CLIP image features with ResNet features (ablation)CLIP-It · SumMe and TVSum generic summarizationF1: ResNet+Transformer -> CLIP-Image+Transformer: SumMe 52.8 -> 53.5, 54.9 -> 55.3, 50.3 -> 51.0; TVSum 65.0 -> 65.5, 67.5 -Hurt
remove classification loss from the supervised objective (ablation)CLIP-It · SumMe and TVSum; standard, augment, and transfer configurationsF1: 54.2 -> 52.5; 56.4 -> 54.7; 51.9 -> 50.0; 66.3 -> 63.0; 69.0 -> 65.7; 65.5 -> 62.8Hurt
add reconstruction loss to classification loss (ablation)CLIP-It · SumMe and TVSum; standard, augment, and transfer configurationsF1: 49.1 -> 53.0; 52.6 -> 55.4; 46.2 -> 50.3; 60.2 -> 64.5; 61.7 -> 66.5; 57.3 -> 63.4Improved
add diversity loss to classification loss (ablation)CLIP-It · SumMe and TVSum; standard, augment, and transfer configurationsF1: 49.1 -> 53.7; 52.6 -> 55.8; 46.2 -> 50.8; 60.2 -> 65.4; 61.7 -> 67.6; 57.3 -> 64.3Improved
combine classification, reconstruction, and diversity losses instead of using classification loss alone (ablation)CLIP-It · SumMe and TVSum; standard, augment, and transfer configurationsF1: 49.1 -> 54.2; 52.6 -> 56.4; 46.2 -> 51.9; 60.2 -> 66.3; 61.7 -> 69.0; 57.3 -> 65.5Improved
replace cross-modal attention with an MLP (ablation)CLIP-It · SumMe and TVSum; standard, augment, and transfer configurationsF1: SumMe 50.6, 51.08, 48.1 -> 54.2, 56.4, 51.9; TVSum 63.0, 65.8, 61.4 -> 66.3, 69.0, 65.5Hurt
add generated captions to CLIP-Image+Transformer (ablation)CLIP-It · QFVS generic summaries (supervised)F1: 63.11 -> 65.92Improved
use ground-truth captions instead of generated captions (ablation)CLIP-It · QFVS generic summariesF1 (Avg): 65.92 -> 75.06 (supervised); 59.28 -> 67.12 (unsupervised)Improved
add generated captions to CLIP Image + Transformer (ablation)CLIP-It · UT Egocentric (supervised)F-Measure: 41.58 -> 44.70Improved
add ground-truth captions to CLIP Image + Transformer (ablation)CLIP-It · TV Episodes (supervised)F-Measure: 47.82 -> 55.34Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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