Skip to content

README.md

Latest commit

 

History

History
80 lines (76 loc) · 2.89 KB

README.md

File metadata and controls

80 lines (76 loc) · 2.89 KB

Multi-Modal Object Classification

This is the official repository for the Multi-Modal Classification.

This challenge focuses on Object Classification utilizing multi-modal data source including RGB, depth, and infrared images. You can visit the official website for more details.

Dataset

In this track, we provide a dataset named MMC (Multi-Modal Object Classification), which comprises 3,000 multi-modal image pairs (2000 for training and 1000 for testing) across 13 classes.

Example

Depth Thermal-IR RGB
Depth Output IR Output RGB Output
Depth Output IR Output RGB Output

Structure

MMC
├── train_2k
│ ├──color
│ │ ├── train_0001.png
│ │ ├── train_0002.png
│ │ ├── ... ...
│ │ ├── train_4000.png
│ ├──depth
│ │ ├── train_0001.png
│ │ ├── train_0002.png
│ │ ├── ... ...
│ │ ├── train_4000.png
│ ├──infrared
│ │ ├── train_0001.png
│ │ ├── train_0002.png
│ │ ├── ... ...
│ │ ├── train_4000.png
│ │ ├── ... ...
| |——train_labels.txt
├── test
│ ├──color
│ │ ├── test_0001.png
│ │ ├── test_0002.png
│ │ ├── ... ...
│ │ ├── test_1000.png
│ ├──depth
│ │ ├── test_0001.png
│ │ ├── test_0002.png
│ │ ├── ... ...
│ │ ├── test_1000.png
│ ├──infrared
│ │ ├── test_0001.png
│ │ ├── test_0002.png
│ │ ├── ... ...
│ │ ├── test_1000.png

Baseline

This code is based on Resnet18.

  • ❗Note!!! The validation set is not provided, you should divide the train set appropriately by yourself to validate during training.
  • We have modified the model to accommodate this multimodal task, while you can also build your own model to accomplish this task

Training

  • Change the root path of the dataset (your path to train_2k)
  • run tain.py Train your own model:
  python train.py --root path_to_train_2k \
    --train_labels train_labels.txt \
    --val_labels val_labels.txt \
    --epochs 80 \
    --eval_period 1 \
    --batch 64 \
    --num_classes 13 \
    --output_file path_to_save_the_log_file_and_the_model \

Testing

Generate the predictions submission.csv for test set: run Inference.py

 python Inference.py --root path_to_test_2k \
   --checkpoint path_to_the_best_model \
   --save path_to_save_the_submissionfile \
   --num_classes 13
  • ❗Note Results submission.csv will be generated automatically and it's the only file you need to submit to the platform for evaluation.