Source-linked AI summary

ActionCLIP: A New Paradigm for Video Action Recognition

Mengmeng Wang, Jiazheng Xing, Yong Liu

arXiv:2109.08472v1cs.CV

TL;DR

Video action recognition traditionally maps labels to fixed categories, limiting transfer to unseen concepts. The paper instead uses video-text matching with semantic label supervision and introduces pre-train, prompt, and fine-tune; ActionCLIP reaches 83.8% top-1 accuracy on Kinetics-400 while supporting zero-shot and few-shot recognition.

  • Problem

    Traditional models predict fixed predefined categories and require additional labeled data to transfer to new unseen concepts, while dataset labels are too succinct for rich language learning.

  • Method

    The paper formulates recognition as multimodal video-text matching and reuses web-pretrained models through prompting followed by end-to-end fine-tuning on target datasets.

  • Results

    ActionCLIP achieves 83.8% top-1 accuracy on Kinetics-400 and reports consistent outperformance of state-of-the-art methods across several public benchmarks.

  • Takeaways & Limitations

    Semantic language supervision broadens action-recognition representations and supports zero-shot/few-shot transfer while reducing the cost of pre-training.

Abstract

from arXiv · show

The canonical approach to video action recognition dictates a neural model to do a classic and standard 1-of-N majority vote task. They are trained to predict a fixed set of predefined categories, limiting their transferable ability on new datasets with unseen concepts. In this paper, we provide a new perspective on action recognition by attaching importance to the semantic information of label texts rather than simply mapping them into numbers. Specifically, we model this task as a video-text matching problem within a multimodal learning framework, which strengthens the video representation with more semantic language supervision and enables our model to do zero-shot action recognition without any further labeled data or parameters requirements. Moreover, to handle the deficiency of label texts and make use of tremendous web data, we propose a new paradigm based on this multimodal learning framework for action recognition, which we dub "pre-train, prompt and fine-tune". This paradigm first learns powerful representations from pre-training on a large amount of web image-text or video-text data. Then it makes the action recognition task to act more like pre-training problems via prompt engineering. Finally, it end-to-end fine-tunes on target datasets to obtain strong performance. We give an instantiation of the new paradigm, ActionCLIP, which not only has superior and flexible zero-shot/few-shot transfer ability but also reaches a top performance on general action recognition task, achieving 83.8% top-1 accuracy on Kinetics-400 with a ViT-B/16 as the backbone. Code is available at https://github.com/sallymmx/ActionCLIP.git

1. Introduction

ActionCLIP reframes video action recognition as multimodal video-text matching, using semantic label information to improve representation and transfer. Its pre-train, prompt, and fine-tune paradigm reuses web-pretrained models while achieving strong benchmark performance.

  • Traditional action recognition predicts fixed predefined categories, limiting transfer to unseen concepts because new datasets require additional labeled training data.
  • The multimodal framework encodes videos and label texts separately, then pulls corresponding representations together through similarity learning.Inference becomes video-text matching rather than 1-of-N classification.
  • The framework strengthens video representations with semantic language supervision and supports zero-shot transfer without further labeled data or parameters.
  • The proposed paradigm reuses large-scale web-pretrained models through prompt design and target-dataset fine-tuning, reducing pre-training cost.It addresses succinct dataset labels and the storage, computation, and experiment costs of end-to-end web-data pre-training.
  • Comprehensive experiments report consistent outperformance of state-of-the-art methods across several public benchmark datasets.

2. Related Works

Related work spans hand-crafted and increasingly engineered video architectures, while vision-text methods motivate treating action recognition as multimodal retrieval. ActionCLIP focuses on incorporating label-text semantics into general action recognition rather than only learning video features or text embeddings.

  • 2.1. Video Action Recognition: Video action recognition progressed from hand-crafted spatio-temporal descriptors to two-stream, 3D CNN, compute-efficient, and transformer-based architectures.
  • 2.1. Video Action Recognition: The paper proposes a pre-trained, prompt, and fine-tune paradigm based on video-text multimodal learning to foreground language modeling of label words.
  • 2.2. Vision-text Multi-modality in Action Recognition: Prior vision-text approaches either learn video representations while adding a classifier or emphasize embedding design with simple text features, leaving general action recognition less addressed.

3. Method

ActionCLIP reframes action recognition as video-text matching rather than unimodal label classification, then combines multimodal learning with pre-training, prompting, and end-to-end fine-tuning. Its framework uses separate video and language encoders, cosine similarity, contrastive KL loss, and temporal visual prompts.

  • 3.1. Multimodal Learning Framework: The framework replaces fixed-label classification with video-text matching, using semantic label information to compare videos with textual categories.Inference selects the label words with the highest similarity score.
  • 3.1. Multimodal Learning Framework: Separate video and language encoders produce spatiotemporal video features and label-text features, which are aligned through symmetric cosine similarities.The encoded representations are compared in both video-to-text and text-to-video directions.
  • 3.1. Multimodal Learning Framework: The method uses a video-text contrastive objective based on symmetric KL divergence because multiple videos in a batch may share one label.This avoids treating similarity learning as a strict 1-in-N classification problem.
  • 3.2. The New Paradigm: The proposed paradigm combines web-data pre-training, textual and visual prompt engineering, and end-to-end fine-tuning on target datasets.Prompting addresses deficient label words while adapting pre-trained representations to action recognition.
  • 3.3. New Paradigm Instantiation Details: ActionCLIP instantiates the paradigm with CLIP, manual textual prompts, and visual prompts that adapt image encoders to spatiotemporal video information.Visual prompt choices include pre-network, in-network, and post-network designs, with temporal modules such as MeanP, Conv1D, LSTM, and Transf.
  • 3.3. New Paradigm Instantiation Details: The complete ActionCLIP network is fine-tuned end-to-end using the video-text contrastive training objective.The architecture includes two single-modal encoders, similarity calculation, and multiple prompt locations.

4. Experiments

Experiments show that ActionCLIP benefits from multimodal supervision, pre-training, suitable prompts, and end-to-end fine-tuning. It also transfers to data-poor settings and achieves strong accuracy while offering efficient configurations.

  • Ablation Experiments: 2.91% top-1 accuracy gains result from exploiting label-text semantics in the multimodal framework.The comparison uses the same backbone, pre-trained weights, and temporal modeling strategy before the final linear layer.
  • Ablation Experiments: 78.36% versus 40.10% shows that proper visual-encoder initialization is especially important for pre-training.Randomly initializing both encoders produces a 41.4% drop from the fully pre-trained model.
  • Ablation Experiments: 2.74% and 5.38% performance decreases occur with Joint and Shift visual prompts, respectively, relative to MeanP.The authors attribute these decreases to catastrophic forgetting caused by changed input patterns or altered pre-trained features.
  • Ablation Experiments: 6.15% lower top-1 accuracy occurs when both encoders are frozen, whereas end-to-end fine-tuning reaches 78.36%.Freezing either pre-trained encoder negatively influences accuracy, supporting dataset-specific end-to-end fine-tuning.
  • Comparison with State-of-the-art Methods: 44.3 mAP is achieved on Charades, demonstrating effectiveness for multi-label video classification.The dataset contains longer-range activities and multiple actions per video.

5. Conclusion

The paper reframes action recognition as video-text multimodal learning and introduces “pre-train, prompt, and fine-tune” to reuse large-scale web-data models. ActionCLIP achieves strong performance on general and zero-shot/few-shot recognition.

  • Action recognition is modeled as a video-text multimodal learning problem that exploits semantic information in label texts.
  • The “pre-train, prompt, and fine-tune” paradigm reuses powerful large-scale web-data pre-trained models while greatly reducing pre-training cost.
  • ActionCLIP delivers superior performance on general and zero-shot/few-shot action recognition.
Loading 2109.08472v1…