Choice 2: Head localization

In this project, you will build a convolutional neural network that estimates the number of individuals in a crowded scene by directly localizing their head positions. Your implementation will follow the approach described in “Rethinking Counting and Localization in Crowds: A Purely Point-Based Framework” by Song et al. (2021). This approach formulates crowd counting as a direct localization problem, predicting head positions through a combination of feature extraction and point regression. The model is trained using a loss function that matches predicted points to ground-truth annotations using the Hungarian algorithm, allowing it to effectively learn to identify and count individuals in dense crowds.

Difficulty	Suggested Tutorials
Medium	Convolutional Networks, Transfer Learning

Requirements: A GPU with at least 8GB of memory is recommended. Training on a CPU is possible but will be significantly slower.

---

Grading

The project will be graded based on the following criteria. Points for each activity are awarded based on quality and completeness (partial credit possible).

Activity	Points (max)
Build a preprocessing pipeline for the ShangaiTech dataset	4
Implement the P2PNet architecture on top of VGG16	4
Implement the loss function with hungarian assignment	2
Train the backend while keeping the frontend frozen	2
Fine-tune the entire network end-to-end	2
Evaluate model performance using MAE	3
Presentation (clarity & demo)	3
Total	0-20