Awesome-Temporal-Action-Detection-Temporal-Action-Proposal-Generation

A reading list that collects research papers on temporal action detection and related tasks. The word **temporal** means “time-based”: these papers are about when actions happen in a video.

What the Project Does

It is a link collection, not software. You cannot run it. It simply organizes hundreds of papers into a browsable index. Each entry links to a research paper. Some entries also link to code repositories, blog posts, or video talks.

The Core Task

Temporal Action Detection (the main topic of this list) means:\n\n> Given an unstructured video, find where in time a human action starts and ends, and what that action is.

Example: A surveillance video might contain “punching” and “running” as separate actions. The goal is to mark the exact start-time and end-time of each.

What the List Covers

1. Temporal Action Proposal Generation

Before detecting actions, many systems first generate proposals (candidate time-windows that might contain an action). This is a separate research line with its own papers.

2. Temporal Action Detection (Supervised)

The mainstream research area. Papers assume training data with full labels (start-time, end-time, action class).

3. Weakly Supervised Detection

Training with less detailed labels. Instead of giving exact times, you only tell the model “somewhere in this video there is a punching action”. The model must figure out the timing itself.

4. Semi-Supervised Detection

Training with a mix of labeled and unlabeled videos. The model uses the labeled ones as anchors and tries to learn from the many unlabeled ones too.

5. Online Action Detection

Detecting actions as the video plays, without waiting for the whole video. Real-time systems need this.

6. Open-Vocabulary Detection

Detecting actions that were never seen during training, described by language. Example: a system trained on “running” and “jumping” must also recognize “skateboarding” from a textual description.

7. Pretrained Model Resources

Five papers that provide pretrained video encoders (trained on large video data) which you can reuse for new detection tasks. Example: BSP (ICCV 2021) and TSP -)

8. ActivityNet Challenge

A link to the 2021 ActivityNet competition (a benchmark competition for activity detection) and a talk-recording link from Alibaba DAMO Academy.

How the List Is Organized

For each topic, papers are grouped by publication year (2016\–2026). This makes it easy to trace how the field evolved.

• 2026 entries are marked as arXiv.
• Older entries (“before”-sections) go back to 2016\–2017.

Types of Links in the List

• Paper links (the majority): arXiv, IEEE Xplore, CVF - Open Access (CVPR/ICCV).
• Code links (some entries): GitHub repositories implementing the paper.
• Zhihu links (a Chinese blogging platform): Chinese-language explanations of certain papers.
• Talk links: e.g. a VALSE talk recording hosted on Baidu Pan (a file-sharing service) with password “7y8g”.

Notable Details

• The list includes both official code (e.g. TensorFlow, PyTorch, PaddlePaddle) and unofficial re-implementations (e.g. code.PyTorch_unofficial). This is stated explicitly in the entries.
• Some entries have multiple code ports (e.g. BSN has both TensorFlow and PyTorch versions).
• The list marks competition results: e.g. RapNet is labeled “2019 ActivityNet task-1 2nd” (second place in that competition).

Audience

Primarily computer-vision researchers (graduate students, PhDs, engineers) looking for papers, code, and baseline implementations. The list does not provide tutorials or hands-on guides for beginners.

Summary

A straightforward, well-organized bibliography. Its value comes from curation (selecting quality papers) and organization (grouping by sub-task and year), not from code or tooling.