This repo is an official implementation of our IJCV paper: Logit Normalization for Long-Tail Object Detection, which was published in 08 January 2024.

Please refer to springer and arxiv for more details about our paper!

Key Idea

Real-world data with skewed distributions poses a serious challenge to existing object detectors. Via adding this LogitNormHead implemented below on the predicted logit vector for calibration, the long-tail bias will get greatly alleviated!

In general, our LogN is training- and tuning-free (i.e. require no extra training and tuning process), model- and label distribution-agnostic (i.e. generalization to different kinds of detectors and datasets), and also plug-and-play (i.e. direct application without any bells and whistles).

class LogitNormHead(Shared2FCBBoxHead):
    def __init__(self, momentum=1e-4, *args, **kwargs):
        super(LogitNormHead, self).__init__(*args, **kwargs)
        self.bn = nn.BatchNorm1d(self.num_classes + 1, eps=1e-05, momentum=momentum, affine=False)
    def get_statistics(self):
        mean_val = self.bn.running_mean
        mean_val[-1] = 0
        std_val = torch.sqrt(torch.clamp(self.bn.running_var, min=1e-11))
        std_val[-1] = 1
        beta = torch.zeros_like(mean_val)
        beta[:-1] = mean_val[:-1].min()
        return mean_val.view(1, -1), std_val.view(1, -1), beta.view(1, -1)
    def forward(self, x):
        cls_score, bbox_pred = super(LogitNormHead, self).forward(x)
        if self.training:
            cls_score = self.bn(cls_score)
            return cls_score, bbox_pred
        else:
            mean_val, std_val, beta = self.get_statistics()
            cls_score = (cls_score - (mean_val - beta)) / std_val
            return cls_score, bbox_pred

Please refer to models/roi_heads/bbox_heads/logit_norm_head.py for more details!

Installation

Please refer to get_started.md for installation.

Prepare LVIS Dataset

for images

LVIS uses same images as COCO's, so you need to donwload COCO dataset at folder ($COCO), and link those train, val under folder lvis($LVIS).

mkdir -p data/lvis
ln -s $COCO/train $LVIS
ln -s $COCO/val $LVIS
ln -s $COCO/test $LVIS

for annotations

Download the annotations from lvis webset

cd $LVIS
mkdir annotations

then places the annotations at folder ($LVIS/annotations)

Finally you will have the file structure like below:

data
  ├── lvis
  |   ├── annotations
  │   │   │   ├── lvis_v1_val.json
  │   │   │   ├── lvis_v1_train.json
  │   ├── train2017
  │   │   ├── 000000004134.png
  │   │   ├── 000000031817.png
  │   │   ├── ......
  │   ├── val2017
  │   ├── test2017

for API

The official lvis-api and mmlvis can lead to some bugs of multiprocess. See issue

So you can install this LVIS API from my modified repo.

pip install git+https://github.com/tztztztztz/lvis-api.git

Pipeline

1. Download the pretrained models from the below urls. You can also train the baselines from scrath, please refer to scripts/baseline.sh;
2. Please refer to scripts/logn.sh. Specifically, for logit-normalized calibration, we are now using an online approach, which involves finetuning the pretrained model for 1 epoch with the LogitNormHead attached for aggregating statistics and then directly perform calibrated evaluation.

Citation

If you use the equalization losses, please cite our papers.

@article{zhao2024logit,
  title={Logit Normalization for Long-tail Object Detection},
  author={Zhao, Liang and Teng, Yao and Wang, Limin},
  journal={International Journal of Computer Vision},
  year={2024},
  publisher={Springer}
}