svip-lab / hrnet-for-fashion-landmark-estimation.pytorch Goto Github PK

Nice work! Did you calculate sigmas (

i just want to know how can we apply the entire repo on another dataset

1. I have followed the issue #5 but unable to produce the better keypoints.
   is there anything else to improve the Visualization part ?

I test the checkpoint, using the command in Readme.
python tools/test.py
--cfg experiments/deepfashion2/hrnet/w48_384x288_adam_lr1e-3.yaml
TEST.MODEL_FILE models/pose_hrnet-w48_384x288-deepfashion2_mAP_0.7017.pth
TEST.USE_GT_BBOX True

I open the debug switch in config file, and the saved image result is very bad. Any show problem?

训练的时候出现以下错误
for catId in catIds}
File "/home/sa/anaconda3/envs/torch1.7/lib/python3.6/site-packages/pycocotools/cocoeval.py", line 229, in computeOks
e = (dx2 + dy2) / vars / (gt['area']+np.spacing(1)) / 2
ValueError: operands could not be broadcast together with shapes (294,) (17,)
@ShenhanQian

how can we get clothes bbox, clothes class, scale and other parameters, if we want to run this repo on test image

Thanks for sharing great job! However, when configuring the dataset, I cannot find bounding box results named 'bbox_result_val.pkl' in the provided OneDrive link. Did I miss anything?

I am aiming to create an inference pipeline for images in the wild. One of the problem that I am facing is to detect bboxs. Can you please upload the model and code for your detector? Otherwise there's also this repo that I found that can be used for this purpose: https://github.com/archana1998/Clothing-Detection

First, thanks for sharing this great work! Here is some issue that I met.

I try to visualize the result by running the script###

python tools/test.py --cfg experiments/deepfashion2/hrnet/w48_384x288_adam_lr1e-3.yaml TEST.MODEL_FILE models/pose_hrnet-w48_384x288-deepfashion2_mAP_0.7017.pth TEST.USE_GT_BBOX True DATASET.MINI_DATASET True TAG 'experiment description' WORKERS 4 TEST.BATCH_SIZE_PER_GPU 8 TRAIN.BATCH_SIZE_PER_GPU 8

the config file is

AUTO_RESUME: false #
CUDNN:
BENCHMARK: true
DETERMINISTIC: false
ENABLED: true
DATA_DIR: ''
GPUS: (1,)
OUTPUT_DIR: 'output'
LOG_DIR: 'log'
WORKERS: 8
PRINT_FREQ: 100
PIN_MEMORY: true

DATASET:
COLOR_RGB: false
DATASET: 'deepfashion2'
DATA_FORMAT: jpg
FLIP: true
NUM_JOINTS_HALF_BODY: 8
PROB_HALF_BODY: 0.3
ROOT: 'data/deepfashion2/'
ROT_FACTOR: 15 #45
SCALE_FACTOR: 0.1 #0.35
TEST_SET: 'validation'
TRAIN_SET: 'train'
MINI_DATASET: True
SELECT_CAT: [1,2,3,4,5,6,7,8,9,10,11,12,13]
MODEL:
INIT_WEIGHTS: true
NAME: pose_hrnet
NUM_JOINTS: 294
PRETRAINED: ''
TARGET_TYPE: gaussian
IMAGE_SIZE:

• 288
• 384
  HEATMAP_SIZE:
• 72
• 96
  SIGMA: 2 # 3
  EXTRA:
  PRETRAINED_LAYERS:
  • 'conv1'
  • 'bn1'
  • 'conv2'
  • 'bn2'
  • 'layer1'
  • 'transition1'
  • 'stage2'
  • 'transition2'
  • 'stage3'
  • 'transition3'
  • 'stage4'
    FINAL_CONV_KERNEL: 1
    STAGE2:
    NUM_MODULES: 1
    NUM_BRANCHES: 2
    BLOCK: BASIC
    NUM_BLOCKS:
    • 4
    • 4
      NUM_CHANNELS:
    • 48
    • 96
      FUSE_METHOD: SUM
      STAGE3:
      NUM_MODULES: 4
      NUM_BRANCHES: 3
      BLOCK: BASIC
      NUM_BLOCKS:
    • 4
    • 4
    • 4
      NUM_CHANNELS:
    • 48
    • 96
    • 192
      FUSE_METHOD: SUM
      STAGE4:
      NUM_MODULES: 3
      NUM_BRANCHES: 4
      BLOCK: BASIC
      NUM_BLOCKS:
    • 4
    • 4
    • 4
    • 4
      NUM_CHANNELS:
    • 48
    • 96
    • 192
    • 384
      FUSE_METHOD: SUM
      LOSS:
      USE_TARGET_WEIGHT: true
      TRAIN:
      BATCH_SIZE_PER_GPU: 8
      SHUFFLE: true
      BEGIN_EPOCH: 0
      END_EPOCH: 210
      OPTIMIZER: adam
      LR: 0.001 #0.001
      LR_FACTOR: 0.1
      LR_STEP:
• 170
• 200
  WD: 0.
  GAMMA1: 0.99
  GAMMA2: 0.0
  MOMENTUM: 0.9
  NESTEROV: false
  TEST:
  BATCH_SIZE_PER_GPU: 8
  COCO_BBOX_FILE: ''
  DEEPFASHION2_BBOX_FILE: ''
  BBOX_THRE: 1.0
  IMAGE_THRE: 0.0 # threshold for detected bbox to be feed into HRNet
  IN_VIS_THRE: 0.2
  MODEL_FILE: ''
  NMS_THRE: 1.0
  OKS_THRE: 0.9 # the lower threshold for a peak point in a heatmap to be kept
  USE_GT_BBOX: true
  FLIP_TEST: true
  POST_PROCESS: true
  SHIFT_HEATMAP: true
  DEBUG:
  DEBUG: True
  SAVE_BATCH_IMAGES_GT: false
  SAVE_BATCH_IMAGES_PRED: false
  SAVE_BATCH_IMAGES_GT_PRED: True
  SAVE_HEATMAPS_GT: false
  SAVE_HEATMAPS_PRED: false

I change the CONFIG parameter to True, however it still does not save any image. The image saving only works when I change the BATH_SIZE_PER_GPU to 1. However, the image-saving function is based on a torch grid, thus result in a very wired visualization since the scale of keypoint and output is different. Could you please try to solve the problem? I am using a single GPU RTX 3080TI with Ubuntu 18.04.

Thanks for great job! And there is a question about the output of pretained model provided.
When I test pretained model, it return a tensor whose shape is [batchsize, 294, height, weight]. Could you give some expanation about the number "294" ?

First, I would like to thank authors for sharing the code. The work inspired me a lot and I‘m trying to followup your method for finding keypoints using my own input images.
However, I found that it is hard to directly modify test.py to meet a situation when the input is not the validation set, which really confused me. Is there a straightforward way to generate landmark with input images not in the deepfashion2 dataset?