Soccer Player Detection: Classification + Conditional Localization

This repo implements an end-to-end vision pipeline for a soccer-playing robot using the Soccer Player Detection dataset. The system learns to understand a cluttered game scene from raw RGB images by solving two complementary tasks:

1. Classification: identify which entity is present (e.g., robot / ball / goalpost).
2. Regression (Localization): predict the normalized center coordinates (x, y) of a requested target object in the full camera frame. :contentReference[oaicite:0]{index=0}

---

What the Project Does

Task 1 — Object Classification (Cropped Inputs)

A lightweight CNN classifies cropped image regions into object categories. The pipeline uses stratified splitting to preserve minority classes and evaluates with accuracy. :contentReference[oaicite:1]{index=1}

Task 2 — Object Localization (Full-Image Regression)

A regression model predicts the target’s center point (cx, cy) in [0, 1] × [0, 1], derived from bounding boxes. To resolve ambiguity in scenes with multiple objects, the localization network is conditional: it takes both the image and a target class ID as inputs, so it can “focus” on the requested object. :contentReference[oaicite:2]{index=2}

---

Key Points

Preventing Data Leakage (Grouped Split)

For localization, splitting is done at the image level (grouped by filename), ensuring objects from the same original frame never appear across train/val/test—preventing background memorization and leakage. :contentReference[oaicite:3]{index=3}

Regression-Safe Online Augmentation

A custom augmentation pipeline applies:

• photometric noise (brightness/contrast)
• horizontal flips with synchronized label updates (flipping cx → 1 - cx)
  This improves stability and reduces overfitting in regression training.

---

Models

Classification CNN

• Input: 128×128 RGB
• 3 Conv blocks (increasing filters), max-pooling
• Dense head + Softmax output :contentReference[oaicite:5]{index=5}

Conditional Localization Network

• Visual branch: CNN feature extractor
• Context branch: class ID → Embedding
• Fusion: concatenate features + embedding
• Output: 2D (x, y) with Sigmoid to keep predictions within image bounds :contentReference[oaicite:6]{index=6}

Optimization uses Adam with a cosine-decay learning-rate schedule and early stopping. :contentReference[oaicite:7]{index=7}

---

Results Summary (from the report)

• Classification: ~98.5% test accuracy with the final optimized CNN. :contentReference[oaicite:8]{index=8}
• Localization: average error reduced from ~28 px (baseline) to ~21.8 px with the conditional architecture (ResNet-style + regularization). :contentReference[oaicite:9]{index=9}
• Per-class localization is hardest on the Robot due to shape/orientation variation. :contentReference[oaicite:10]{index=10}

---

Repo Contents (typical)

• notebooks / scripts for training classification and regression models
• dataset loader with:
  • filename reconciliation
  • stratified split for classification
  • grouped split for regression
• augmentation pipeline that keeps labels consistent with transforms

---

Notes / Limitations

While classification is near-perfect, localization error can still be significant for fine control, and heavy occlusion cases may remain challenging. Future improvements suggested include moving to a single-shot detector (e.g., YOLO-style) and temporal smoothing (e.g., Kalman filtering). :contentReference[oaicite:11]{index=11}