Classification Matters: Improving Video Action Detection with Class-Specific Attention

Paper | Project Page

This repository is the official implementation of "Classification Matters: Improving Video Action Detection with Class-Specific Attention" (ECCV 2024 Oral)

Classification Matters: Improving Video Action Detection with Class-Specific Attention
Jinsung Lee¹, Taeoh Kim², Inwoong Lee², Minho Shim², Dongyoon Wee², Minsu Cho¹, Suha Kwak¹
POSTECH¹, NAVER Cloud²
accepted to ECCV 2024 as an oral presentation

Detection Result	Classification Attention Map 1	Classification Attention Map 2
$~~~~~~$ $~~~~~~$	(talk to)	(listen to)
$~~~~~~$ $~~~~~~$	(answer phone)	$~~~~$ (listen to) $~~~~$

Installation

The code works on

• Ubuntu 20.04
• CUDA 11.7.0
• CUDNN 8.0.5
• NVIDIA A100 / V100

Install followings,

• Python: 3.8.10
• GCC 9.4.0
• PyTorch: 2.0.0

and run the installation commands below:

pip install -r requirements.txt
cd ops
pip install .

Data Preparation

Refer here for AVA preparation. We use updated annotations (v2.2) of AVA. Download annotation assets and place it outside the project folder (../assets).

Refer here for UCF101-24 preparation.

Refer here for JHMDB51-21 preparation.

Running commands

Our model is trained in two steps: (following TubeR)

First, it is trained from scratch. Second, it is trained again, but it uses the transformer weights acquired from the first stage.

For convenience, we provide the pre-trained transformer weights of the first stage that are used to train the model.

Evaluation Code

## Evaluate

# AVA 2.2
python3 evaluate.py --pretrained_path={path to the model to evalute} --config-file=./configuration/AVA22_CSN_152.yaml
python3 evaluate.py --pretrained_path={path to the model to evalute} --config-file=./configuration/AVA22_ViT-B.yaml
python3 evaluate.py --pretrained_path={path to the model to evalute} --config-file=./configuration/AVA22_ViT-B_v2.yaml

# UCF
python3 evaluate.py --pretrained_path={path to the model to evalute} --config-file=./configuration/UCF_ViT-B.yaml

# JHMDB (split 0)
python3 evaluate.py --pretrained_path={path to the model to evalute} --config-file=./configuration/JHMDB_ViT-B.yaml --split 0

Model Zoo

Backbone .pth files are the same ones from here (CSN152) and here (ViT-B). We offer this link for the aggregated backbone .pth files.

Dataset	Backbone	Backbone pretrained on	transformer weights	f-mAP	v-mAP	config	checkpoint
AVA 2.2	CSN-152	K400	link	33.5	-	config	link
AVA 2.2	ViT-B	K400	link	32.9	-	config	link
AVA 2.2	ViT-B	K400, K710	link	38.4	-	config	link
UCF	ViT-B	K400	link	85.9	61.7	config	link
JHMDB (split 0)	ViT-B	K400	link	88.1	90.6	config	link

Acknowledgments

Our code is based on DETR, DAB-DETR, Deformable-DETR, and TubeR. If you use our model, please consider citing them as well.

License

Class Query
Copyright (c) 2024-present NAVER Cloud Corp.
CC BY-NC 4.0 (https://creativecommons.org/licenses/by-nc/4.0/)