The evaluation will consist of a two-phase process focused on assessing the precision of the reconstructed 3D models in terms of both **shape (3D structure)** and **portion size (volume)**. Results will be compared against reference models of the same food items captured using a 3D scanner.

### Phase I: Volume Accuracy (measured in mL)

In Phase I, we will use the **Mean Absolute Percentage Error (MAPE)** metric to evaluate the accuracy of predicted portion sizes. Only **volume accuracy** will be considered during this stage.

### Phase II: Shape Accuracy (Top Teams) (measured in mm^3)

The top-ranking teams (3-5, depends on the total number of teams) from Phase I will proceed to Phase II, where they will submit **complete 3D mesh files** for each food item. We will use the **L1 Chamfer Distance** metric to evaluate the **shape accuracy** of the models. Our evaluation protocol follows the Chamfer Distance metric used in **DTU** benchmark [[1]](#dtu).

The **final winner** will be determined by combining scores from both phases for all private food objects. Each team will receive a ranking (1–5) in each phase, and the final score will be a weighted sum:

- **Phase I (Volume Accuracy):** 55% weight  
- **Phase II (Shape Accuracy):** 45% weight

---

## 📦 Phase II Evaluation (Top Teams)

### 1. Upload Models  
Submit all private **reconstructed 3D models** as `.obj` files, each named according to the corresponding **object ID**. Upload the files to a downloadable location (e.g., Google Drive, Dropbox).

### 2. Send the Download Link  
Email the **download link** and your **team ID** to us by **May 13rd, 11:59 PM AOE**.

- **Email:** `metafood.cvpr@gmail.com`  
- For support, post in the **Kaggle discussion forum**.

### 3. Model Verification  
We will verify your submitted models against your final Phase I predictions and visually inspect them for compliance.  
_Models composed of primitive shapes (e.g., spheres or cubes) will be disqualified._

### 4. Ground Truth Models & Evaluation Script  
We will release the **ground truth 3D models** and **Chamfer Distance evaluation script** on Kaggle by **May 13rd, 11:59 PM AOE**.

### 5. Align Models  
You must align your models to the ground truth by preparing a **transformation matrix** for each object.  
We will compute your Chamfer Distance score as follows:

- Randomly sample 16384 points from each transformed point cloud, **10 times**  
- Compute the **L1-Chamfer Distance** for each sample  
- Average the 10 scores to obtain your **final Chamfer Distance score**

### 6. Submit Transformation Matrices  
Email us your transformation matrices by **May 16th, 11:59 PM AOE**.

---

### 🏆 Winner Announcement  
The winning team will be announced on **May 18th**.


## Submission File
For each id in the test set, you must reconstruct a 3D model from the provided images and report the volume of the model. The file should contain a header and have the following format:

    id,predicted
    2,0
    5,0
    6,0
    etc.

<a name="dtu"></a>
### 📚 Reference

[1] Aanæs, Henrik, et al. "*Large-scale data for multiple-view stereopsis.*" *International Journal of Computer Vision*, 120 (2016): 153–168.