Every Pixel Matters: Center-aware Feature Alignment for Domain Adaptive Object Detector

[Project Page] [PDF]

This project hosts the code for the implementation of Every Pixel Matters: Center-aware Feature Alignment for Domain Adaptive Object Detector (ECCV 2020).

Introduction

A domain adaptive object detector aims to adapt itself to unseen domains that may contain variations of object appearance, viewpoints or backgrounds. Most existing methods adopt feature alignment either on the image level or instance level. However, image-level alignment on global features may tangle foreground/background pixels at the same time, while instance-level alignment using proposals may suffer from the background noise.

Different from existing solutions, we propose a domain adaptation framework that accounts for each pixel via predicting pixel-wise objectness and centerness. Specifically, the proposed method carries out center-aware alignment by paying more attention to foreground pixels, hence achieving better adaptation across domains. To better align features across domains, we develop a center-aware alignment method that allows the alignment process.

We demonstrate our method on numerous adaptation settings with extensive experimental results and show favorable performance against existing state-of-the-art algorithms.

Installation

Check INSTALL.md for installation instructions.

The implementation of our anchor-free detector is heavily based on FCOS (#f0a9731).

Dataset

All details of dataset construction can be found in Sec 4.2 of our paper.

We construct the training and testing set by three following settings:

• Cityscapes -> Foggy Cityscapes
  • Download Cityscapes and Foggy Cityscapes dataset from the link. Particularly, we use leftImg8bit_trainvaltest.zip for Cityscapes and leftImg8bit_trainvaltest_foggy.zip for Foggy Cityscapes.
  • Download and extract the converted annotation from the following links: Cityscapes and Foggy Cityscapes (COCO format).
  • Extract the training sets from leftImg8bit_trainvaltest.zip, then move the folder leftImg8bit/train/ to Cityscapes/leftImg8bit/ directory.
  • Extract the training and validation set from leftImg8bit_trainvaltest_foggy.zip, then move the folder leftImg8bit_foggy/train/ and leftImg8bit_foggy/val/ to Cityscapes/leftImg8bit_foggy/ directory.
• Sim10k -> Cityscapes (class car only)
  • Download Sim10k dataset and Cityscapes dataset from the following links: Sim10k and Cityscapes. Particularly, we use repro_10k_images.tgz and repro_10k_annotations.tgz for Sim10k and leftImg8bit_trainvaltest.zip for Cityscapes.
  • Download and extract the converted annotation from the following links: Sim10k (VOC format) and Cityscapes (COCO format).
  • Extract the training set from repro_10k_images.tgz and repro_10k_annotations.tgz, then move all images under VOC2012/JPEGImages/ to Sim10k/JPEGImages/ directory and move all annotations under VOC2012/Annotations/ to Sim10k/Annotations/.
  • Extract the training and validation set from leftImg8bit_trainvaltest.zip, then move the folder leftImg8bit/train/ and leftImg8bit/val/ to Cityscapes/leftImg8bit/ directory.
• KITTI -> Cityscapes (class car only)
  • Download KITTI dataset and Cityscapes dataset from the following links: KITTI and Cityscapes. Particularly, we use data_object_image_2.zip for KITTI and leftImg8bit_trainvaltest.zip for Cityscapes.
  • Download and extract the converted annotation from the following links: KITTI (VOC format) and Cityscapes (COCO format).
  • Extract the training set from data_object_image_2.zip, then move all images under training/image_2/ to KITTI/JPEGImages/ directory.
  • Extract the training and validation set from leftImg8bit_trainvaltest.zip, then move the folder leftImg8bit/train/ and leftImg8bit/val/ to Cityscapes/leftImg8bit/ directory.

After the preparation, the dataset should be stored as follows:

[DATASET_PATH]
└─ Cityscapes
   └─ cocoAnnotations
   └─ leftImg8bit
      └─ train
      └─ val
   └─ leftImg8bit_foggy
      └─ train
      └─ val
└─ KITTI
   └─ Annotations
   └─ ImageSets
   └─ JPEGImages
└─ Sim10k
   └─ Annotations
   └─ ImageSets
   └─ JPEGImages

Format and Path

Before training, please checked paths_catalog.py and enter the correct data path for:

• DATA_DIR
• cityscapes_train_cocostyle, cityscapes_foggy_train_cocostyle and cityscapes_foggy_val_cocostyle (for Cityscapes -> Foggy Cityscapes).
• sim10k_trainval_caronly, cityscapes_train_caronly_cocostyle and cityscapes_val_caronly_cocostyle (for Sim10k -> Cityscapes).
• kitti_train_caronly, cityscapes_train_caronly_cocostyle and cityscapes_val_caronly_cocostyle (for KITTI -> Cityscapes).

For example, if the datasets have been stored as the way we mentioned, the paths should be set as follows:

• Dataset directory (In L8):

  DATA_DIR = [DATASET_PATH]
  

• Train and validation set directory for each dataset:

  "cityscapes_train_cocostyle": {
      "img_dir": "Cityscapes/leftImg8bit/train",
      "ann_file": "Cityscapes/cocoAnnotations/cityscapes_train_cocostyle.json"
  },
  "cityscapes_train_caronly_cocostyle": {
      "img_dir": "Cityscapes/leftImg8bit/train",
      "ann_file": "Cityscapes/cocoAnnotations/cityscapes_train_caronly_cocostyle.json"
  },
  "cityscapes_val_caronly_cocostyle": {
      "img_dir": "Cityscapes/leftImg8bit/val",
      "ann_file": "Cityscapes/cocoAnnotations/cityscapes_val_caronly_cocostyle.json"
  },
  "cityscapes_foggy_train_cocostyle": {
      "img_dir": "Cityscapes/leftImg8bit_foggy/train",
      "ann_file": "Cityscapes/cocoAnnotations/cityscapes_foggy_train_cocostyle.json"
  },
  "cityscapes_foggy_val_cocostyle": {
      "img_dir": "Cityscapes/leftImg8bit_foggy/val",
      "ann_file": "Cityscapes/cocoAnnotations/cityscapes_foggy_val_cocostyle.json"
  },
  "sim10k_trainval_caronly": {
    "data_dir": "Sim10k",
      "split": "trainval10k_caronly"
  },
  "kitti_train_caronly": {
      "data_dir": "KITTI",
      "split": "train_caronly"
  },
  

(Optional) Format Conversion

If you want to construct the dataset and convert data format manually, here are some useful links:

• yuhuayc/da-faster-rcnn
• krumo/Detectron-DA-Faster-RCNN

Training

To reproduce our experimental result, we recommend training the model by following steps.

Let's take Cityscapes -> Foggy Cityscapes as an example.

1. Pre-training with only GA module

Run the bash files directly:

• Using VGG-16 as backbone with 4 GPUs

  bash ./scripts/train_ga_vgg_cs.sh
  

• Using ResNet-101 as backbone with 4 GPUs

  bash ./scripts/train_ga_resnet_cs.sh
  

• (Optional) Using VGG-16 as backbone with single GPU

  bash ./scripts/single_gpu/train_ga_vgg_cs_single_gpu.sh
  

or type the bash commands:

• Using VGG-16 as backbone with 4 GPUs

  python -m torch.distributed.launch \
      --nproc_per_node=4 \
      --master_port=$((RANDOM + 10000)) \
      tools/train_net_da.py \
      --config-file ./configs/da_ga_cityscapes_VGG_16_FPN_4x.yaml
  

• Using ResNet-101 as backbone with 4 GPUs

  python -m torch.distributed.launch \
      --nproc_per_node=4 \
      --master_port=$((RANDOM + 10000)) \
      tools/train_net_da.py \
      --config-file ./configs/da_ga_cityscapes_R_101_FPN_4x.yaml
  

• (Optional) Using VGG-16 as backbone with single GPU

  python tools/train_net_da.py \
      --config-file ./configs/da_ga_cityscapes_VGG_16_FPN_4x.yaml \
      SOLVER.MAX_ITER 80000 \
      SOLVER.IMS_PER_BATCH 4
  

2. Training with both GA and CA module

First, set the MODEL.WEIGHT as the path of pre-trained weight in L5 of the config file (example).

Next, the model can be trained by the following commands:

Run the bash files directly:

• Using VGG-16 as backbone with 4 GPUs

  bash ./scripts/train_ga_ca_vgg_cs.sh
  

• Using ResNet-101 as backbone with 4 GPUs

  bash ./scripts/train_ga_ca_resnet_cs.sh
  

• (Optional) Using VGG-16 as backbone with single GPU

  bash ./scripts/single_gpu/train_ga_ca_vgg_cs_single_gpu.sh
  

or type the bash commands:

• Using VGG-16 as backbone with 4 GPUs

  python -m torch.distributed.launch \
      --nproc_per_node=4 \
      --master_port=$((RANDOM + 10000)) \
      tools/train_net_da.py \
      --config-file ./configs/da_ga_ca_cityscapes_VGG_16_FPN_4x.yaml
  

• Using ResNet-101 as backbone with 4 GPUs

  python -m torch.distributed.launch \
      --nproc_per_node=4 \
      --master_port=$((RANDOM + 10000)) \
      tools/train_net_da.py \
      --config-file ./configs/da_ga_ca_cityscapes_R_101_FPN_4x.yaml
  

• (Optional) Using VGG-16 as backbone with single GPU

  python tools/train_net_da.py \
      --config-file ./configs/da_ga_ca_cityscapes_VGG_16_FPN_4x.yaml \
      SOLVER.MAX_ITER 80000 \
      SOLVER.IMS_PER_BATCH 4
  

Note that the optimizer and scheduler will not be loaded from the pre-trained weight in the default setting. You can set load_opt_sch as True in train_net_da.py to change the setting.

Evaluation

The trained model can be evaluated by the following command.

python tools/test_net.py \
	--config-file [CONFIG_PATH] \
	MODEL.WEIGHT [WEIGHT_PATH] \
	TEST.IMS_PER_BATCH 4

• [CONFIG_PATH]: Path of config file
• [WEIGHT_PATH]: Path of model weight for evaluation

For example, the following command evaluates the model weight vgg_cs.pth for Cityscapes -> Foggy Cityscapes using VGG-16 backbone.

python tools/test_net.py \
	--config-file configs/da_ga_ca_cityscapes_VGG_16_FPN_4x.yaml \
	MODEL.WEIGHT "vgg_cs.pth" \
	TEST.IMS_PER_BATCH 4

Note that the commands for evaluation are completely derived from FCOS (#f0a9731).

Please see here for more details.

Result

We provide the experimental results and model weights in this section.

*Since the original model weight for KITTI dataset is inaccessible for now, we re-run the experiment and provide a similar (and even better) result in the table.

Note that we use 4 GPUs for faster training. For fair comparison, we also report the results using only one GPU.

Environments

• Hardware

  • 4 NVIDIA 1080 Ti GPUs

• Software

  • PyTorch 1.3.0
  • Torchvision 0.2.1
  • CUDA 10.2

Citations

Please consider citing our paper in your publications if the project helps your research.

@inproceedings{hsu2020epm,
  title     = {Every Pixel Matters: Center-aware Feature Alignment for Domain Adaptive Object Detector},
  author    = {Cheng-Chun Hsu, Yi-Hsuan Tsai, Yen-Yu Lin, Ming-Hsuan Yang},
  booktitle = {European Conference on Computer Vision},
  year      = {2020}
}