pipixin321 / hr-pro

Official implementation of HR-Pro(AAAI2024)

Hi there 👋, I'm Huaxin Zhang

I am a Master of HUST (Huazhong University of Science and Technology), supervised by Nong Sang.

🔭 Reseach-wise, I mainly focus on:

• Multi-modal Large Language Models
• Video Understanding, more specifically, Weakly-supervised Temporal Action Localization (WSTAL) & Weakly-suervised Video Anomaly Detection (WSVAD).

😄 I am open to:

• A internship/job/PhD offer with computer vision/multimodal LLM research and engineering.

📫 Contact me by:

• Email: [email protected]

💬 News:

• 2024-07-01: We release our code and model of "Holmes-VAD: Towards Unbiased and Explainable Video Anomaly Detection via Multi-modal LLM".[project page]
• 2024-06-10: We release our code and model of "Arcana: Improving Multi-modal Large Language Model through Boosting Vision Capabilities".[project page]
• 2024-01-29: I start my internship in Baidu VIS, to do some research on Multi-modal Large Language Model (MLLM).
• 2023-12-09: One paper about point supervised temporal action localization is accepted on AAAI 2024.

---

hr-pro's Issues

custom_data 特征提取时选用的fps, stride, num_frames

在训练自己的数据时，stride, num_frames应该设置为多少？ 另外在代码里面看到hard code编码，所以这里stride都要使用16吗？
self.t_factor_point = self.args.frames_per_sec / (vid_fps * 16)

其他三个数据集的训练参数配置信息

你好，麻烦能提供其他三个数据集的训练参数配置文件么，感谢🙏

你好，请问能公布下其它三个数据集的yaml文件吗

About ActivityNet1.3 Point_labels

Thank for you great work!
However, I'm wondering about the ActivityNet1.3 point_labels file:
Why is its' start_frame 'and' stop_frame 'not integer but float type, and does 'point' provide the frame index ?

How to find the dataset features and annotations used in GTEA, BEOID, and ActivityNet 1.3?

Thank you for the significant advancements you've brought. I am a beginner in this fields. Could you please guide me on how to find the dataset features and annotations used in GTEA, BEOID, and ActivityNet 1.3?