Indoor

Detection in Indoor Scenes

Provided task is split into 3 stages:

• Analyse: Data Visualization and Explanation
• Make Splits: Preprocessing data for training
• Training: Model training and validation

First two stages are provided in notebooks folder
Code for the last part is stored in indoor folder

---

Code Running

• Recreate train/val split using notebooks
• Install egg pip3 install -e .
• Install dependencies pip3 install -r requirements.txt

Trainig: python3 indoor/training/train.py --kwargs...
Visualization: python3 indoor/project_utils/visualize.py --kwargs...

Trained model is available at Google Drive

---

List of functionality

Implemented
Model Training
Predicts visualization
mAP and mAR metrics
per class mAP metrics

Not Implemented but could be useful
Share info between nearby frames
Sequence business metric*
Modulated deformable convs**
box_score thrs for each class
ablation study for params, archs etc.

*Metric that shows model quality on each sequence of frames **Should work fine for this task

---

Stack & Arch & Params

• DL lib: Pytorch x Torchvision

• MlOps & Tracking: ClearML

• Detector: Faster-RCNN based 2 stage detector

• Backbone: ResNeST 50 (ResNet 50 with split Attention)

• Batch size: 4

• Epochs: 50

• Base Lr: 0.0005

• Pretrain: None

• Optimizer: Ranger (Lookahead + RAdam)

• Lr_scheduler: ReduceOnPlateo with reloading best checkpoint*

*That's my favourite, always SOTA

---

Results

Best mAP@50: 97.2

mAP per class

Class	mAP
fireextinguisher	0.75
chair	0.79
exit	0.79
clock	0.76
trashbin	0.61
screen	0.69
printer	0.79

Good examples

Bad examples

FP for different classes

Next frame example

At evaluation should aggregate predictions from nearby frames because of cases like this

Summary

• Got pretty good working model but there are many ways to increase final quality
• Statements mentioned in "Not Implemented but could be useful" are crucial but not included in this work