Signate segmentation
My solution for AI Edge Contest (Segmentation Division) - 22 place from 402 participants with 0.652 IoU.
Data
The training data for segmentation models consist of 2,243 images, test data - 649. Images have resolution of 1936 x 1216. Categories utilized for submission evaluation (but totally in data - 20 categories):
- car
- pedestrian
- signal
- lane.
Notebook with EDA.
Model
I used Unet with different pretrained backbones. Train dataset was splitted on 7 folds statified by route category and time of day for validation. Notebook with model's errors analysis.
Methods
- Train on crops (256px, 512px, 768px); predict on full, predict on crops, predict on crops with intersection.
- Train on resized (50%, 75% of original size); predict on full, predict on resized + bilinear upsampling.
- Train on full size; predict on full size - best score.
- Train on all categories (not only on 4); predict 4 categories.
What is implemented
- learning rate scheduling
- early stopping
- different backbones (ResNet family + WideResnet)
- different loss functions (bce, jaccard, focal)
- a lot more (see input args of
train.py)
What didn't work
- In-place activated batchnorm - doesn't reduce memory usage too much
- fp16 training - it's really free a lot GPU memory and you can use bigger batch size,
but model didn't converge to good results (yep, I tried
dynamic_loss_scale=True)
TODO
- strong image augmentations
- one cycle learning rate
- snapshot ensembling + folds ensembling
- SE, scSE blocks in decoder
- TTA
- LinkNet
- pseudo labeling
- more experiments with loss combinations
Training
Best model (resnet 34) was trained on full size images with batch_size=3 on Titan V CEO Edition with 32 Gb memory (thanks to Scitator).
Training was on all 20 categories with bce loss for 20 epochs (12 hours).
Prediction visualization
Top image - raw, middle - ground true, bottom - model prediction.
Model trained on 4 categories
Model trained on all categories
