Skip to content

Latest commit

 

History

History
63 lines (39 loc) · 3.38 KB

Backbone.md

File metadata and controls

63 lines (39 loc) · 3.38 KB
Raw

Backbone

To be continued...

Backbone K Modality Flip FrameAP @0.5 VideoAP @0.2 | @0.5 | @0.75 | 0.5:0.95 FPS Download
DLA-34 7 RGB ✓ 73.14 78.81 | 51.02 | 27.05 | 26.51 29 model
✗ 72.05 78.23 | 50.77 | 26.10 | 26.16 53 ⬆️
ResNet-18 7 RGB ✓ 70.36 76.31 | 50.03 | 25.66 | 25.72 50 model
✗ 68.63 76.70 | 49.31 | 24.63 | 25.11 85 ⬆️

(All experiments validate on UCF101-24 test split with coco pretrain)

Online FPS tests on a single NVIDIA TITAN XP with --batch_size 1.


Train RGB K=7 on UCF for ResNet-18

Firstly, download coco pretrained ResNet-18 model from this.

We get this pretrained model from Centernet, which adds three up-convolutional layers to obtain a higher-resolution output.

Please move this pretrained model to ${MOC_ROOT}/experiment/modelzoo


Then, run

python3 train.py --K 7 --exp_id Train_K7_rgb_coco_resnet18 --rgb_model $PATH_TO_SAVE_MODEL --batch_size 128 --master_batch 16 --lr 5e-4 --gpus 0,1,2,3,4,5,6,7 --num_workers 16 --num_epochs 10 --lr_step 5,8 --save_all --arch resnet_18

Don't forget to add --arch resnet_18.


Evaluate RGB K=7 on UCF for ResNet-18

Download our result model from this. Then run:

python3 det.py --task normal --K 7 --gpus 0,1,2,3,4,5,6,7 --batch_size 128 --master_batch 16 --num_workers 16 --rgb_model ../experiment/result_model/$PATH_TO_RGB_MODEL --inference_dir $INFERENCE_DIR --flip_test --arch resnet_18
python3 det.py --task normal --K 7 --gpus 0 --batch_size 1 --master_batch 1 --num_workers 0 --rgb_model ../experiment/result_model/$PATH_TO_RGB_MODEL --inference_dir $INFERENCE_DIR --flip_test --arch resnet_18

python3 ACT.py --task frameAP --K 7 --th 0.5 --inference_dir $INFERENCE_DIR

python3 ACT.py --task BuildTubes --K 7 --inference_dir $INFERENCE_DIR

python3 ACT.py --task videoAP --K 7 --th 0.2 --inference_dir $INFERENCE_DIR
python3 ACT.py --task videoAP --K 7 --th 0.5 --inference_dir $INFERENCE_DIR
python3 ACT.py --task videoAP --K 7 --th 0.75 --inference_dir $INFERENCE_DIR
python3 ACT.py --task videoAP_all --K 7 --inference_dir $INFERENCE_DIR

Don't forget to add --arch resnet_18.


Bash File

We also provide bash file for training. Please refer train_ucf_k7_resnet18.sh.