[Approximate Supervised Object Distance Estimation] Problem with Dataset

[Yihang Feng]

Dear MaCVi, i can't find the download link for Approximate Supervised Object Distance Estimation dataset.

[Benjamin Kiefer]

Sorry, the README going to get updated today. 

[Benjamin Kiefer]

We are facing some issues with the uploading the code which should be resolved in the next few days. In the meantime, please see this subset that already gives you an understanding of how the data is set up:

https://drive.google.com/drive/folders/1sF7DX3ekSH4xZS5D5sqOuB7c2SiaWPyi?usp=sharing

It's YOLO format for the labels except for the last column, which is the metric distance in meters. 

[Benjamin Kiefer]

Category information:

 0=boat, 1=other (almost exclusively buoys, maybe some poles, debris or similar as well)

Zitierten Text anzeigen

[Matej Fabijanic]

Hello,

do you have an estimation on when the complete distance dataset will be available to download?

Regards,
Matej

[Benjamin Kiefer]

Hi,

sorry for the delay! We found some inaccuracies in the dataset and had to create a new one.

There will be minor changes in the category (likely buoy only as opposed to boat and buoy).

We'll try to upload today or tomorrow, including a starter kit.

Thanks!

[Benjamin Kiefer]

Please find an update online

https://macvi.org/workshop/macvi25/challenges/usv_dist

On Monday, October 28, 2024 at 2:52:18 PM UTC+1 Matej Fabijanic wrote:

[ZQ Z]

Hello, organizer, can only one model have the ability to detect and calculate distance at the same time? Can two models be used for target detection and distance calculation, and then the two models can be merged?

Benjamin Kiefer 在 2024年11月1日 星期五凌晨3:27:37 [UTC+8] 的信中寫道：

[Benjamin Kiefer]

Hi!

Yes, as long as it fits the structure of the required evaluation protocol. I'd suggest you create such a model, export it to ONNX and test whether the evaluation runs locally. The test script testscript_onnx.py you can find in the starter code:

https://github.com/Ben93kie/YOLOv7-DL23/tree/distance_network

Have you tried merging your two models somehow? Maybe like this:

https://chatgpt.com/share/67360992-0d88-8001-b2ca-933a0cfe5c5d

Let me know how it is going.

-Ben

[ZQ Z]

Is the code for the online evaluation similar to the test script testscript_onnx.py? I need more details because the output of the onnx model will affect the online evaluation.

Benjamin Kiefer 在 2024年11月14日 星期四晚上10:31:20 [UTC+8] 的信中寫道：

[Benjamin Kiefer]

Yes, it's very similar. If it works for you local script, it will work for the online evaluation. 

In any case, let me know how I can assist. 

[ZQ Z]

Hello, organizer, can you provide relevant papers to describe how you modified yolov7 to be able to predict distance? Thank you

Benjamin Kiefer 在 2024年11月18日 星期一下午6:41:45 [UTC+8] 的信中寫道：

[Benjamin Kiefer]

It's currently under review. We might put it on arxiv earlier. But, it's not a crazy method. It's essentially just another predicted parameter per anchor. 
Let me know if I can answer any specific questions! 

[ZQ Z]

Is he using linear regression to predict the distance? Can you point out the key code? Thank you

Benjamin Kiefer 在 2024年11月20日 星期三下午5:21:02 [UTC+8] 的信中寫道：

[Benjamin Kiefer]

It's learning-based. Does that make sense? 

[Marten Kre]

Hi,

you can find the adaptions to the model architecture in the models/yolo.py file.

We modified the Detect / IDetect Layers, which are the final layers of the YOLOv7 Net w.r.t. Object Detection.

These Layers provide the output vector for each anchor consiting of (x_center, y_center, width, height, objectness, class_probs).

We extended this vector with the distance estimate -> (x_center, y_center, width, height, objectness, class_probs, dist_estimate).

During training and inference, sigmoid is applied to the distance estimate (line 164 for inference, for training see loss.py).

We use the normalized distance (according to normalization strategy in hyperparameters file) for training, so there is no need to rescale the output.

During inference however, we strive to provide a metric estimate, thus we need to rescale the value (inverse of normalizaion), see line 167.

You can inspect the commit history of the repository to see all the changes to the architecture.

The most important ones are to:

• Model Architecture (i just explained that part)
• Loss Function
• Data Loader
• Train / Test Scripts
  

[ZQ Z]

Hello, organizer. The online test has an error, but the offline test can run normally. Why is this? Thank you

Marten Kre 在 2024年11月20日 星期三晚上11:07:01 [UTC+8] 的信中寫道：

[Benjamin Kiefer]

Hm, we're just integrating the evaluation code for another challenge track and did restart the webserver a couple of times.
It could be that it is related to that.

Could you reupload the onnx?

If it crashes again, pls let me know and I'll take a closer look.

[ZQ Z]

I uploaded it three times and it's still the same

ben93...@gmail.com 在 2024年12月1日 星期日晚上9:23:49 [UTC+8] 的信中寫道：

[Benjamin Kiefer]

thx, looking into it.. give me a few mins

[Benjamin Kiefer]

Hm, your latest upload was some time ago and in that time period where we restarted the server a couple of times. I'm uploading your model under our name for testing to see whether it goes through. You can try the same so it counts towards your uploads. In a few hours, we'll have an answer then. If it still fails, we'll need to look into the actual model.

[ZQ Z]

It seems to work fine, thanks

ben93...@gmail.com 在 2024年12月1日 星期日晚上9:51:01 [UTC+8] 的信中寫道：

[ZQ Z]

Hello, organizer, there is an error when uploading the model. Have you restarted the server or is it my problem?

ZQ Z 在 2024年12月1日 星期日晚上10:27:55 [UTC+8] 的信中寫道：

[Benjamin Kiefer]

Hi, did you previous submissions work? Then, it's likely related to image restoration challenge webpage updates.. Sorry for that!

[ZQ Z]

The one I uploaded previously was invalid.

ben93...@gmail.com 在 2024年12月2日 星期一晚上7:24:01 [UTC+8] 的信中寫道：

[Marten Kre]

Hi, 

i checked your previous uploads and it seems that the issue was indeed caused by server restarts, since the interpreter did not exit with an error message.

At the moment multiple models are executed on the server, so it might take even longer than expected for the evaluation to finish.

I tested the failed upload #13193 with the testscript_onnx locally and this also worked without issues.

I case future uploads still fail please don't hesitate to notify us.

[Benjamin Kiefer]

Are you still perceiving issues?

[ZQ Z]

No, thank you

ben93...@gmail.com 在 2024年12月3日 星期二上午8:29:00 [UTC+8] 的信中寫道：