A specialist is needed for training a model for object detection in architectural drawings.

Budget: 27000 UAH Deadline: 21 days

We have a practically ready approach for object detection in drawings, which can be quickly adapted to your classes and brought to a working JSON output. ))

The estimate is 90,000 UAH and 21 days for the first stage without backend.

This stage includes an audit of 20-30 typical drawings, preparation of the annotation format, training of the basic model, a Python script for execution, an example JSON, README, and a report on precision, recall, and mAP.

Here’s the nuance... 90%+ precision and recall can only be promised after checking the dataset and the quality of the annotations.

If the annotations are not yet available, they need to be accounted for separately or a semi-automatic annotation process should be established - otherwise, the model will start making beautiful mistakes, and this is not the theater where tickets are needed.

- Approach - bbox detection for electrical, lighting, furniture, and architectural elements.
- We will separately consider rotations of symbols at 45, 90, and 180 degrees.
- Result - trained model, execution script, JSON, README, metrics, and recommendations for further training.

Clarifications:
- Do you already have annotated drawings with bbox and classes, or only raw files?
- The files are mainly PDF, DWG, PNG, or JPG.

Relevant examples from Ingello:
- https://business.ingello.com/fractal - close in building complex AI automation around business processes.
- https://business.ingello.com/vorfahr - an example of an AI approach where stable applied logic is important, not just an experiment.

The main profile of Ingello for the exchange - https://systems-fl.ingello.com/ua

We can start with a small stage of data verification and prototyping, so as not to raise expectations until the real quality of the drawings is visible (:. Expand

Budget: 22500 UAH Deadline: 12 days

Good day. I understand the task: to train or adapt a detection model that finds elements (electrical, lighting, furniture, architectural symbols) in architectural drawings, outlines them with a bbox, determines the class, and returns the result in JSON with a count. The target quality is 90%+ for precision and recall.

I see the approach as follows: a modern detection model is taken for fine-tuning on your 500-800 drawings. First, there is a data labeling stage according to your list of classes, then iterative training with metric monitoring on a test sample. It is also important to consider rotations of symbols at 45, 90, and 180 degrees and different graphic styles, and some classes like specification tables and frames should be taken not only through detection but also in relation to the structure of the sheet.

Honestly about the realism of the target: 90% across all classes at once is ambitious, as rare lighting symbols and small markings learn worse with a small number of examples. Therefore, it is usually done step by step: first, we confidently take frequent classes, then we refine the model on weaker ones to achieve the desired quality.

To assess the scope: are your 500-800 drawings already labeled for these classes, or does labeling also need to be done within the project? This has the most impact on timelines.

Budget: 20000 UAH Deadline: 10 days

Good day!

I am interested in your project. I have experience working with Computer Vision, Object Detection, and Python (particularly in the defense sector), so I understand the specifics of automatic analysis of technical drawings and can implement a "turnkey" solution.

### How I see the project implementation

**1. Data analysis and preparation**

* quality check of drawings;
* unification of formats;
* recommendations for optimal labeling;
* preparation of train/validation/test samples.

**2. Model training**
A modern architecture (YOLO / RT-DETR or similar, depending on testing results) will be used for detection, which works well with a large number of classes and provides high inference speed.

During training, the following will be applied:

* rotation augmentation (45°, 90°, 180°, and arbitrary angles);
* scaling;
* line thickness variation;
* noise and contrast;
* other augmentation methods to enhance the model's generalization ability.

This will allow the model to work stably with drawings of different styles, languages, and conditional symbols.

### Detection

The model will find:

* electrical elements;
* lighting;
* furniture;
* plumbing;
* architectural elements;
* specifications and service blocks.

For each object, the following will be returned:

```json
{
"class": "socket",
"label": "220V Socket",
"bbox": [x, y, width, height],
"confidence": 0.94,
"source": "model",
"modifier": "3",
"quantity": 3
}
```

### What you will receive

✅ Trained model

✅ Python inference script

✅ JSON output

✅ README for execution

✅ Metrics:

* Precision
* Recall
* mAP

✅ Testing examples on new drawings

✅ Recommendations for further labeling and retraining of the model.

### Quality

The main goal will be not just to train the model on one type of drawing, but to make it as robust as possible to different styles of presentation, symbol rotations, scales, and differences between design projects.

Before the final handover, I will conduct testing on a separate test sample and provide a brief report with actual quality metrics.

I would be happy to discuss the project details, review examples of drawings, and propose an optimal approach to achieve 90%+ accuracy. Expand

Budget: 5000 UAH Deadline: 1 day

Good day. I am ready to complete this project; I have extensive experience in developing various applications.

Budget: 27000 UAH Deadline: 60 days

Good day. I created a neural network for myself to detect text presence in images - I did it using batches of 32x32 - if you're interested, I can show the results in private messages. So I think I can create such a neural network - it can also be done using larger batches. I have a question about the parameters modifier and quantity - please show an example when there will be values there.
60000 UAH / 60 days

Budget: 25000 UAH Deadline: 14 days

We have experience in developing computer vision systems for specific technical documents. We implement this through fine-tuning the YOLOv8 architecture on your dataset, followed by optimization for complex geometric shapes in drawings. This will ensure high detection accuracy even at large scales. We are ready to discuss the details and approach to data annotation.