Thread by @ai_fast_track

✨ Training Object Detection Models Tips & Tricks ✨

🧊 Data Labeling:

📌 Avoid adding low quality data: it confuses your model

📌 Prevent data leak

📌 Dataset size: Smaller size if using pretrained models. Bigger size if training from scratch

...

📌 Use prototypical (representative) data for each class

📌 Identify incorrect classes

📌 Identify ambiguous labelled images

📌 Balance your data distribution

📌 Train from scratch if your dataset is different from the COCO dataset

📌 When your model stops improving, you might add more data to move the needle:

📌 Use Soft-Labelling: label new data with pretrained models => Free labels

📌 Use Self-Training: label new data with your model you are training. Add newly labelled data to your training dataset. Loop

🧊 Modeling

📌 Use larger models. They outperform smaller ones

📌 Use smaller models 😀 when training small dataset

📌 Use Focal Loss for the classification head (RetinaNet , EfficientDet, ...)

📌 Use GIoU Loss for the regression head (box location): Need a separate post 😄

📌 Trained with ImageNet isn't always effective

📌 YOLO-ReT paper showed that using the whole pretrained backbone harms model performance

📌 Use Backbone Truncation (like YOLO-ReT): Only 60% of the backbone is initialised with pretrained weights.

🧊 Anchor Boxes

📌 Use anchor boxes with size/ratio close to target boxes

📌 Make sure you know how your anchor boxes look like

📌 Use some anchor-free OD models (e.g., VFNet): No anchor boxes. Some perform even better than anchor-based models

🧊 Data augmentation

📌 Oversample images with small boxes

📌 Use transforms close to your use case

📌 Use Copy & Paste boxes data augmentation

📌 Use mosaic data augmentation

📌 Some suggest using heavy data augmentation at the beginning of training, and light data augmentation at the end

🧊 Training

📌 Don’t worry too much about overfitting at the beginning of project: Adding more data/ data augmentation should mitigate that

📌 Train for a longer period (more epochs): As long as your loss is decreasing

📌 In transfer learning, when freezing NN layers, you should leave BatchNorm layers as trainable

📌 Train using progressive resizing

📌 Use discriminative learning rate. Low learning rate for backbones, higher learning rate for the head

📌 Use the recommended LR scheduler

🧊 Inference

📌 Use the same image size as the one you train your model with

📌 With high resolution images, apply inference on patches/slices and then stitch them together: e.g., Slicing-Aided Hyper Inference

📌 Don’t forget to put the model on evaluation mode (eval_mode): It automatically disables Dropout, BatchNorm, and backpropagation.

🎉This is my longest thread since I joined🐦

If you like this kind of content, follow @ai_fast_track for more OD / CV demystified content in your feed

🙏If you could give the thread a quick retweet, it would help others discover this content! Thanks!

https://twitter.com/ai_fast_track/status/1458493806362378244