AbdurRahman011/csrnet-indian-metro-crowd-density

CSRNet Fine-tuned — Indian Metro Crowd Density Estimation

Fine-tuned CSRNet for real-time crowd density estimation on Indian metro platforms.

MAE = 12.36 on held-out Indian metro test set. 95.6% more accurate than YOLOv8 on the same data.

Model Details

• Architecture: CSRNet (VGG-16 frontend + dilated conv backend, dilation rate r=2)
• Training data: ShanghaiTech Part A + B (700 images) + Custom Indian Metro (88 images)
• Custom dataset: 5,030 head-point annotations across 111 images from Delhi Metro (Rajiv Chowk), Hyderabad Metro (Ameerpet), and Mumbai Central
• Best epoch: 14 | MAE: 11.30 | MSE: 13.99
• Inference: < 0.5 seconds on CPU

Results

Approach	MAE	Density Map	Verdict
CNN Classifier (ResNet-18)	~55	No	Failed
YOLOv8 (pretrained COCO)	283.23	No	Failed
CSRNet Pretrained	~50	Yes	Baseline
CSRNet Fine-tuned (this)	12.36	Yes	Best

Usage

import torch
from huggingface_hub import hf_hub_download
from model import CSRNet

weights_path = hf_hub_download(
    repo_id  = "AbdurRahman011/csrnet-indian-metro-crowd-density",
    filename = "csrnet_v3_best.pth"
)

model = CSRNet()
model.load_state_dict(torch.load(weights_path, map_location="cpu"))
model.eval()

Author

Abdur Rahman Qasim — B.Tech CSE 2025–26
Methodist College of Engineering and Technology, Hyderabad
Guide: Dr. Shivani Yadao