The official PyTorch code for NeurIPS 2021 ML4AD Paper, "Does Thermal data make the detection systems more reliable?"

Last update: Nov 02, 2022

Related tags

Overview

MultiModal-Collaborative (MMC) Learning Framework for integrating RGB and Thermal spectral modalities

This is the official code for NeurIPS 2021 Machine Learning for Autonomous Driving Workshop Paper, "Does Thermal data make the detection systems more reliable?" by Shruthi Gowda, Elahe Arani and Bahram Zonooz.

Methodology

Architecture

Detection Head : SSD
Detection Backbone : Resnet (CNN-based) or DEiT (Transformer-based)

MMC framework

MMC framework has multiple versions

KD.ENABLE: True
KD.ENABLE_DML: True

1. MMC (Base Version) : Det Loss + DML Loss 
    KD.DISTILL_TYPE : KL, AT, L2, L2B
    KL (KL divergence), AT (Attention loss), L2 (L2 norm at head layer), L2B (L2 norm of backbone features)
   
2. MMC v1 (Reconstruction) : Det Loss + DML Loss + Recon Loss
    KD.AUX_RECON = True
    KD.AUX_RECON_MODE = "normal"

3. MMC v2 (Cross Reconstruction) : Det Loss + DML Loss + Cross Recon Loss
    KD.AUX_RECON = True
    KD.AUX_RECON_MODE = "cross"

We also try other techniques for comparison

Fusion
1. Input Fusion
    KD.CONCAT_INPUT
2. Feature Fusion
    KD.CONCAT_FEATURES
    CONCAT_LAYERS

Installation

You can prepare the environment using:

pip install -r requirements.txt

You can build the project using the following script:

./build {conda_env_name}

Datasets

Two datasets "FLIR" and "KAIST" are used in this repo

FLIR : https://www.flir.eu/oem/adas/adas-dataset-form/

KAIST : https://soonminhwang.github.io/rgbt-ped-detection/

Running

Train

There are 2 networks, one receiving RGB images and one receiving thermal images. Both require different config files.

python train.py --config-file <thermal-config-file> --teacher-config-file <rgb-config-file>

Test

For evaluation only one network is used - the first network (RGB or Teacher network)

python test.py --config-file <config-file> --ckpt <model_final.pth>

Model Checkpoints

Cite Our Work

License

This project is licensed under the terms of the MIT license.

The official PyTorch code for NeurIPS 2021 ML4AD Paper, "Does Thermal data make the detection systems more reliable?"

Related tags

Overview

MultiModal-Collaborative (MMC) Learning Framework for integrating RGB and Thermal spectral modalities

Methodology

Architecture

MMC framework

Installation

Datasets

Running

Train

Test

Model Checkpoints

Cite Our Work

License

Owner

NeurAI

HiFT: Hierarchical Feature Transformer for Aerial Tracking (ICCV2021)

Implementation of "Distribution Alignment: A Unified Framework for Long-tail Visual Recognition"(CVPR 2021)

PyTorch wrappers for using your model in audacity!

This repository is an implementation of paper : Improving the Training of Graph Neural Networks with Consistency Regularization

Official implementation of the ICCV 2021 paper "Conditional DETR for Fast Training Convergence".

Constraint-based geometry sketcher for blender

meProp: Sparsified Back Propagation for Accelerated Deep Learning

Code to reproduce the experiments from our NeurIPS 2021 paper " The Limitations of Large Width in Neural Networks: A Deep Gaussian Process Perspective"

Deep Two-View Structure-from-Motion Revisited

Awesome Weak-Shot Learning

GeDML is an easy-to-use generalized deep metric learning library

Deepfake Scanner by Deepware.

Neural style in TensorFlow! 🎨

Official implementation of "One-Shot Voice Conversion with Weight Adaptive Instance Normalization".

Official pytorch implementation of the AAAI 2021 paper Semantic Grouping Network for Video Captioning

CVAT is free, online, interactive video and image annotation tool for computer vision

Bayesian Neural Networks in PyTorch

scikit-learn inspired API for CRFsuite

implement of SwiftNet:Real-time Video Object Segmentation