Our #ECCV2026 @eccvconf paper Gen2Balance: Generative Balancing for Long-Tailed Video Action Recognition now on ArXiv w/ code&data&models https://prajwalgatti.github.io/gen2balance/ Can we fill imbalanced action recognition datasets with generative images to outperform long-tail SOTA? Yes we can 🧵
No Digg Deeper questions have been answered for this story yet.
#1499@eccvconf We introduce automatic pipeline that synthesises videos conditioned on diverse, class-faithful prompts generated by a multimodal LLM. This avoids class ambiguities - e.g. 'robot-dancing' in Kinetics might be misunderstood as a robot dancing rather than dancing like a robot! 2/N
Our #ECCV2026 @eccvconf paper Gen2Balance: Generative Balancing for Long-Tailed Video Action Recognition now on ArXiv w/ code&data&models https://prajwalgatti.github.io/gen2balance/ Can we fill imbalanced action recognition datasets with generative images to outperform long-tail SOTA? Yes we can 🧵
@eccvconf Gen2Balance extends to truly rare actions, with sig gains +31% We release all 140K generated videos to enable reproducibility. Code&models already published. Work led by @prajwal_gatti @bristolcs w great collab @FabianCabaH &Simon Jenni @AdobeResearch. https://arxiv.org/abs/2606.22416 N/N
@eccvconf You can check class-specific gains &how generated data helps per example on our interactive webpage https://prajwalgatti.github.io/gen2balance/index.html#explorer 4/N
@eccvconf We employ a two-stage training (real+synt then real) and train with a class-balanced loss, using margins derived from real-data frequencies. Gen2Balance outperforms strongest long-tail baselines by +6.7% on K100, +5.5% on UCF few-shot classes, +31.9% on actions from RareAct 3/N
@eccvconf We introduce automatic pipeline that synthesises videos conditioned on diverse, class-faithful prompts generated by a multimodal LLM. This avoids class ambiguities - e.g. 'robot-dancing' in Kinetics might be misunderstood as a robot dancing rather than dancing like a robot! 2/N
@eccvconf You can check class-specific gains &how generated data helps per example on our interactive webpage https://prajwalgatti.github.io/gen2balance/index.html#explorer 4/N
@eccvconf We employ a two-stage training (real+synt then real) and train with a class-balanced loss, using margins derived from real-data frequencies. Gen2Balance outperforms strongest long-tail baselines by +6.7% on K100, +5.5% on UCF few-shot classes, +31.9% on actions from RareAct 3/N
