AI Paper Atlas

Research map / Video / Video generation

GePSAn: Generative Procedure Step Anticipation in Cooking Videos

Mohamed A. Abdelsalam, Samrudhdhi B. Rangrej, Isma Hadji, Nikita Dvornik, Konstantinos G. Derpanis, Afsaneh Fazly

ICCV 2023 · Video › Video generation · approach: Compositional text-to-video diffusion · 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
use a generative CVAE recipe encoder with five sampled predictions instead of a single predictionGePSAn · YouCookII video, unseen split, finetunedBLEU4: 4.8 -> 7.9Improved
replace baseline with GePSAn under multiple-prediction evaluationbaseline with nucleus sampling · YouCookII video, unseen split, finetunedingredient recall (ING): 27.8 -> 31.6Improved
add auxiliary MSE loss L_aux to train CVAE (ablation)GePSAn · Recipe1M+ (text), multiple predictionsingredient recall (ING): 29.4 -> 37.2Improved
add next-step prediction loss L_pred (ablation)GePSAn · Recipe1M+ (text), multiple predictionsBLEU4: 7.7 -> 10.7Improved
add sentence reconstruction loss L_rec (ablation)GePSAn · Recipe1M+ (text)ingredient recall (ING), multiple: 36.6 -> 37.2No clear effect
keep the KL divergence term in the objective (ablation)GePSAn · Recipe1M+ (text)training stabilityImproved
transfer text-pretrained model to video zero-shot via aligned UniVL encodersGePSAn (M) · YouCookII video, unseen splitingredient recall (ING): 30.0 zero-shot vs 27.8 finetuned baselineImproved
finetune on video data after text pretrainingGePSAn (M) · YouCookII video, unseen splitBLEU4: 3.7 -> 7.9Improved
use frozen UniVL text encoder for text pretraining instead of learning from scratchGePSAn · Recipe1M+ (text), single predictionVERB recall: 29.4 -> 28.5No clear effect
train from scratch on YouCookII without text pretraining (ablation)GePSAn · YouCookIIperformanceHurt
add cross-modal contrastive distillation to the baselinebaseline · YouCookII video, seen split, multipleBLEU4: 5.9 -> 6.8Improved
zero-shot dataset transfer in single-prediction setting vs baselineGePSAn (S) · YouCookII text, unseen splitingredient recall (ING): 22.9 -> 20.0Hurt

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