Paper | Pretrained models | OpenImages download

Official PyTorch Implementation

Emanuel Ben-Baruch, Tal Ridnik, Itamar Friedman, Avi Ben-Cohen, Nadav Zamir, Asaf Noy, Lihi Zelnik-Manor
DAMO Academy, Alibaba Group

Abstract

Large-scale multi-label classification datasets are commonly, and perhaps inevitably, partially annotated. That is, only a small subset of labels are annotated per sample. Different methods for handling the missing labels induce different properties on the model and impact its accuracy. In this work, we analyze the partial labeling problem, then propose a solution based on two key ideas. First, un-annotated labels should be treated selectively according to two probability quantities: the class distribution in the overall dataset and the specific label likelihood for a given data sample. We propose to estimate the class distribution using a dedicated temporary model, and we show its improved efficiency over a naive estimation computed using the dataset's partial annotations. Second, during the training of the target model, we emphasize the contribution of annotated labels over originally un-annotated labels by using a dedicated asymmetric loss. Experiments conducted on three partially labeled datasets, OpenImages, LVIS, and simulated-COCO, demonstrate the effectiveness of our approach. Specifically, with our novel selective approach, we achieve state-of-the-art results on OpenImages dataset.

We provide direct and convenient access for the OpenImages (V6) dataset. This will enable a common and reproducible baseline for benchmarking and future research. See further details here.

An overview of our approach is summarized in the following figure:

Our loss consists of a selective approach that adjusts the training mode for each class individually and a partial asymmetric loss.

An implementation of the Class-aware Selective Loss (CSL) can be found here.

• class PartialSelectiveLoss(nn.Module)

We provide models pretrained on the OpenImages dataset with different partial training-modes and architectures:

We provide inference code, that demonstrates how to load the model, pre-process an image and do inference. An example run of OpenImages model (after downloading the relevant model):

python infer.py  \
--dataset_type=OpenImages \
--model_name=tresnet_m \
--model_path=./models_local/mtresnet_opim_86.72.pth \
--pic_path=./pics/10162266293_c7634cbda9_o.jpg \
--input_size=224

Training code is provided in train.py. Also, code for simulating partial annotation for the MS-COCO dataset is available (coco_simulation). In particular, two "partial" simulation schemes are implemented: fix-per-class(FPC) and random-per-sample (RPS).

• FPC: For each class, we randomly sample a fixed number of positive annotations and the same number of negative annotations. The rest of the annotations are dropped.
• RPS: We omit each annotation with probability p.

Pretrained weights using the ImageNet-21k dataset can be found here: link
Pretrained weights using the ImageNet-1k dataset can be found here: link

Example of training with RPS simulation:

--data=/datasets/COCO/COCO_2014
--model-path=models/pretrain/mtresnet_21k
--gamma_pos=0
--gamma_neg=1
--gamma_unann=4
--simulate_partial_type=rps
--simulate_partial_param=0.5
--partial_loss_mode=selective
--likelihood_topk=5
--prior_threshold=0.5
--prior_path=./outputs/priors/prior_fpc_1000.csv

Example of training with FPC simulation:

--data=/mnt/datasets/COCO/COCO_2014
--model-path=models/pretrain/mtresnet_21k
--gamma_pos=0
--gamma_neg=3
--gamma_unann=4
--simulate_partial_type=fpc
--simulate_partial_param=1000
--partial_loss_mode=selective
--likelihood_topk=5
--prior_threshold=0.5
--prior_path=./outputs/priors/prior_fpc_1000.csv

The training code contains also the procedure for estimating the class distribution from the data. Our approach enables us to rank the classes based on predictions of a temporary model trained using the Ignore mode. link

@misc{benbaruch2021multilabel,
      title={Multi-label Classification with Partial Annotations using Class-aware Selective Loss}, 
      author={Emanuel Ben-Baruch and Tal Ridnik and Itamar Friedman and Avi Ben-Cohen and Nadav Zamir and Asaf Noy and Lihi Zelnik-Manor},
      year={2021},
      eprint={2110.10955},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Several images from OpenImages dataset are used in this project. Some components of this code implementation are adapted from the repository https://github.com/Alibaba-MIIL/ASL.