2025 At-Home Round

📡 Radar

Detect humans in radar heatmaps — image classification on data that isn't photos.

CVCNNWeird data as images★★☆☆ core CNN reps

📜The task, in plain English

Given radar range–azimuth heatmaps (static and dynamic maps from a radar sensor), decide whether a human is present. The data looks nothing like a photo — it's a 2-D intensity map — but that doesn't matter: it's a grid of numbers, so it's an image to a CNN.

InputRadar heatmap tensors (static + dynamic channels)
OutputHuman present: yes/no (classification)
Really testsRunning an image-classification pipeline on unfamiliar data; channel handling; not panicking at non-photo inputs

🔧The baseline you're given

A small CNN (or simple classifier) trained on the raw heatmaps with a standard loop. Limitations: minimal architecture, no normalization/augmentation strategy tuned to radar data, static and dynamic maps possibly underused as complementary channels.

🚀Baseline vs. solution

⚠️ The baseline (what you are given)
  • Small CNN on raw maps
  • Channels handled naively
  • Little augmentation, default preprocessing
✅ The winning approach
  • Stack static + dynamic maps as channels of one input so the network sees both views at once
  • Normalize per channel (radar intensities have very different scales than photos)
  • Transfer learning: adapt a pretrained ResNet18 — replace the first conv (accept N channels instead of 3) and final layer (2 classes)
  • Augment thoughtfully: flips may be fine, but rotations can break range/azimuth semantics — think before augmenting non-photo data

🧒Explain it like I'm brand new

Key mental unlock: a CNN doesn't know what a photo is. It sees a grid of numbers with local structure. Spectrograms, radar maps, satellite bands, even game boards — if nearby values relate to each other, CNN recipes apply. IOAI loves this trick: Speech Detector (spectrograms), Weather (satellite), Night Watch 2026 (audio-as-image). Learn the ResNet transfer recipe once, reuse forever.

💬The Gemma 4 playthrough (2000-token limit)

Chat 1 · Adapt ResNet to N-channel input
YOU
torchvision resnet18 (weights already local). My input is a 2-channel 64x64 tensor, 2 output classes. Give ONLY the code to modify the first conv layer to accept 2 channels and the fc layer for 2 classes. Max 12 lines.
GEMMA 4
import torch.nn as nn
from torchvision.models import resnet18

m = resnet18()
m.load_state_dict(torch.load('resnet18.pth'))
m.conv1 = nn.Conv2d(2, 64, kernel_size=7, stride=2, padding=3, bias=False)
m.fc = nn.Linear(m.fc.in_features, 2)
Check: out = m(torch.randn(2,2,64,64)); print(out.shape) → [2,2]. Note pretrained conv1 weights are lost when replaced — fine here; you keep the pretrained deeper layers.

🎯Takeaways & what Day 1 might do with this

  • Pattern family: "weird data as images". The 2025 Contest Day 1 shipped a Radar v2 extension — new twist, same pipeline. If your at-home pipeline is clean, Day 1 is a re-run with edits.
← 🦎 Chameleon