reproduce procession

[Cht2924]

Hello, I want to reproduce your code, but I don't know what the process is, it is to first fuse the two modal data under the Preprocess file to form a third modal image, and then train the three modal images in Distillation. But the RGBT-CC annotation is a json version, can you tell me how to reproduce it?

[HenryCilence]

Hello, and thank you for your attention.

As you mentioned, the preprocessing and distillation stages serve to generate the initial broker modality and subsequently train the broker modality generator. Notably, these two stages do not require dataset annotations.

During the fine-tuning stage, annotations are stored in .npy files, which are converted from their .json counterparts. In these .npy files:

The first two columns represent the position of each annotation point.

The third column records the distances between each annotation point and its nearest neighbors.

For reference, you can check the format used in Bayesian Crowd Counting.

In the RGBT-CC dataset, the .json files contain both the position of each annotation and the total count. These can be easily reorganized into the required .npy format following the structure mentioned above.

If needed, please leave your email address and I can share the .npy files I have generated.

[Cht2924]

Thank you very much for your answer，My email is 1774573360@qq.com，Thanks again！！！

[Li-Yijun-lyj]

你好，可以分享一下生成的.npy文件吗？非常感谢，我的邮箱是1679520291@qq.com

[HenryCilence]

你好，已发送

[Li-Yijun-lyj]

好的，感谢！

[Cht2924]

你好，想请问一下密度计数这一领域，是否将batchsize设为1，得到的模型效果更好？

[HenryCilence]

你好，对于本方法而言，将batchsize设为1得到的模型效果更好；但就整个计数领域而言也不一定，比如在UCF-QNRF数据集中，对于Bayesian Loss和DM-Count方法，batchsize设置成1，8或者更大，对性能影响不大。

[Cht2924]

好的，感谢你的回答！

…

------------------ 原始邮件 ------------------ 发件人: "HenryCilence/Broker-Modality-Crowd-Counting" ***@***.***>; 发送时间: 2025年6月20日(星期五) 上午10:24 ***@***.***>; ***@***.******@***.***>; 主题: Re: [HenryCilence/Broker-Modality-Crowd-Counting] reproduce procession (Issue #4) HenryCilence left a comment (HenryCilence/Broker-Modality-Crowd-Counting#4) 你好，对于本方法而言，将batchsize设为1得到的模型效果更好；但就整个计数领域而言也不一定，比如在UCF-QNRF数据集中，对于Bayesian Loss和DM-Count方法，batchsize设置成1，8或者更大，对性能影响不大。 — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>