P—06 Case study · Research
MiPa — Mixed Patch Object Detection
Role
Computer Vision Researcher
Status
completed
Started
Mar 2026
Last activity
Mar 2026
Stack
8 technologies
MiPa — Mixed Patch Object Detection
Python · PyTorch · torchvision
Overview
A research critique and experimental reproduction of the WACV 2025 paper 'Mixed Patch Visible-Infrared Modality Agnostic Object Detection.' Implemented and compared 5 mixing-ratio distributions (Uniform, Beta(2,2), Beta(0.5,0.5), Truncated Uniform, Gaussian) across CNN (ResNet50-FPN) and ViT (Swin-Tiny) backbones for pedestrian detection on the LLVIP dataset. Key finding: Beta(2,2) consistently outperforms the paper's default Uniform(0,1) distribution — achieving +1.24% RGB AP50 improvement on CNN backbones. Note: Due to Kaggle GPU session constraints, CNN backbone results were partially saved; the ViT backbone experiments have complete result logs with consistent improvement across all mixing distributions.
Outcomes
RGB AP50 Gain
Distributions
Backbones
Avg AP50
Impact
- 01Critiqued and reproduced WACV 2025 paper — identified that Uniform(0,1) mixing ratio is suboptimal for modality-agnostic detection
- 02Proved Beta(2,2) distribution outperforms paper default by +1.24% RGB AP50 on CNN backbones with 6-epoch training
- 03Implemented MiPa patch mixing on both CNN (ResNet50-FPN) and ViT (Swin-Tiny) backbones using vectorized tensor operations
- 04Achieved 86.40% Avg AP50 on LLVIP pedestrian detection with only 4K training samples and 12.8-minute training time
Features
01
5 Distribution Comparison
Uniform, Beta(2,2), Beta(0.5,0.5), Truncated Uniform, and Gaussian mixing ratios compared systematically
02
Dual Backbone Validation
CNN (ResNet50-FPN with Faster R-CNN) and ViT (Swin-Tiny) — cross-architecture validation of all mixing strategies
03
WACV 2025 Reproduction
Independent critique and experimental reproduction finding Uniform(0,1) suboptimal vs Beta(2,2)
04
Beta(2,2) Proven Best
+1.24% RGB AP50 improvement over paper default on CNN; consistent gains across all ViT distributions tested
05
LLVIP Dataset
12,025 paired visible-thermal (RGB-IR) pedestrian images — 4K training samples with zero-shot evaluation
06
Kaggle GPU Pipeline
Vectorized tensor operations for efficient patch mixing — 12.8-minute training on single GPU (T4 x2)
Architecture
MiPa (Mixed Patch) mixing layer inserted at the input stage: visible (RGB) and infrared (IR) images are divided into patches, then mixed using a sampling distribution to create modality-agnostic training samples. Distribution parameter p controls the ratio of visible-to-infrared patches. The mixing layer is agnostic to the downstream detector — tested with both Faster R-CNN (ResNet50-FPN) and ViT-based (Swin-Tiny) detection heads. Experiments run on Kaggle GPUs (Tesla T4 x 2) with mixed-precision training over 6 epochs.
Stack
Development log
· Started
Extra comparisons and analysis beyond the MiPa paper
Experimented with additional comparisons and analysis for MiPa. The paper misses detailed reasoning of many choices and my experiments showed improved performance using variable learning rate with cosine decay and other adaptive lr techniques. Also better performance using different sampling distributions compared to just using uniform like in the paper.