DeepFashion Fashion Detection using Tensorflow Object Detection API

Requires Python 3. Tested on Python 3.7.6 in macOS 10.13.6 and 10.14.5, and on Python 3.5.2 in Ubuntu 16.04.6 LTS. Does not work on Python >= 3.8.

Installation

Clone Tensorflow Models Repository and put inside tf/ folder (tf/models/)
Follow the setup for Object Detection API. Remember to specify tensorflow==1.14 or tensorflow-gpu==1.14, tensorflow 2.0 does not work for this API.
pip install --requirement requirements.txt (for extra dependencies used in this repo)

Recommended Directory Structure

+accuracy_confusionMatrix
  -accuracy_confusionMatrix.py
+data
  -label_map file
  -train TFRecord file
  -test TFRecord file
  -eval TFRecord file
+docs
+generate_tfrecord
  -deep_fashion_to_tfrecord.py
+inference
  -Object_detection_image.py
  -Object_detection_webcam.py
+models
  + <model name>
    -pipeline config file
    +train
    +eval
+tf
  +models (tensorflow models repository)
-requirements.txt

Preparing Environment Variables

export PYTHONPATH=$PYTHONPATH:<path to project directory>/tf/models/research:<path to project directory>/tf/models/research/slim

Remember to run this in every terminal session prior to doing anything else. Or you can put this in your .bashrc (.bash_profile for macOS) file for persistency.

Dataset Preparation

# From the project root directory
DATASET_PATH={Path to DeepFashion project dataset with Anno, Eval and Img directories e.g. /home/user/deepfashion/}
OUTPUT_PATH={Path to output TFRecord e.g. data/val.record}
CATEGORIES={broad or fine, broad FOR top, bottom or full only, fine FOR categories.}
EVALUATION_STATUS={train, val or Test}
LABEL_MAP_PATH={Path to label map proto, e.g. data/deepfashion_label_map_fine.pbtxt.}

python generate_tfrecord/deep_fashion_to_tfrecord.py \
    --dataset_path ${DATASET_PATH} \
    --output_path ${OUTPUT_PATH} \
    --categories ${CATEGORIES} \
    --evaluation_status ${EVALUATION_STATUS} \
    --label_map_path ${LABEL_MAP_PATH}

Training and Evaluation

Training and evaluation using the sample configs:

# From the tensorflow/models/research/ directory
cd tf/models/research/

PIPELINE_CONFIG_PATH={path to pipeline config file e.g. models/<model name>/pipeline.config}
MODEL_DIR={path to model directory e.g. models/<model name>/}
NUM_TRAIN_STEPS={60000 was used FOR SsdResnet50, 3000000 was used FOR fasterRcnnInceptionV2}
SAMPLE_1_OF_N_EVAL_EXAMPLES=50

python object_detection/model_main.py \
    --pipeline_config_path=${PIPELINE_CONFIG_PATH} \
    --model_dir=${MODEL_DIR} \
    --num_train_steps=${NUM_TRAIN_STEPS} \
    --sample_1_of_n_eval_examples=$SAMPLE_1_OF_N_EVAL_EXAMPLES \
    --alsologtostderr

Running Tensorboard:

MODEL_DIR={path to model directory e.g. models/<model name>/}
tensorboard --logdir=${MODEL_DIR}

Exporting a Trained Model

# From tensorflow/models/research/
cd tf/models/research/

INPUT_TYPE=image_tensor
PIPELINE_CONFIG_PATH={path to pipeline config file e.g. models/<model name>/pipeline.config}
TRAINED_CKPT_PREFIX={path to model.ckpt e.g. models/<model name>/model.ckpt-<CHECKPOINT_NUMBER>}
EXPORT_DIR={path to folder that will be used For Export e.g. models/<model name>/inference_graph }

python object_detection/export_inference_graph.py \
    --input_type=${INPUT_TYPE} \
    --pipeline_config_path=${PIPELINE_CONFIG_PATH} \
    --trained_checkpoint_prefix=${TRAINED_CKPT_PREFIX} \
    --output_directory=${EXPORT_DIR}

Generating Top-N Accuracy and Confusion Matrix

# From the project root directory

CKPT_PATH={Path to frozen detection graph .pb file, Which contains the model that is used e.g. models/<model name>/inference_graph/frozen_inference_graph.pb}
LABEL_MAP_PATH={Path to label map proto, e.g. data/deepfashion_label_map_fine.pbtxt.}
NUM_CLASSESS={Number of classes the labels have. e.g. 50 FOR DeepFashion}
TFRECORD_PATH={Should use evaluation Set .e.g data/eval.record}
SKIP_EVERY={Same as sample_1_of_n_eval_examples, skip every n images during evaluation. e.g. 50 was used In the paper}
TOP={Top N accuracy needed. e.g. 1 or 3 were used In the paper}
MINIMUM_SCORE={Threshold FOR minimum score that is considered a True positive. e.g. 0.3}
OUTPUT_RESULT={Path to a text file containing all the results}
CONFUSION_TOPN={True or False FOR whether the confusion matrix should follow top N results.}

python accuracy_confusionMatrix/accuracy_confusionMatrix.py \
    --ckpt_path ${CKPT_PATH} \
    --label_map_path ${LABEL_MAP_PATH} \
    --num_classes ${NUM_CLASSESS} \
    --tfrecord_path ${TFRECORD_PATH} \
    --skip_every ${SKIP_EVERY} \
    --top ${TOP} \
    --minimumScore ${MINIMUM_SCORE} \
    --outputResult ${OUTPUT_RESULT} \
    --confusion_topn ${CONFUSION_TOPN}

Inferencing on Webcam

# From the project root directory

CKPT_PATH={Path to frozen detection graph .pb file, Which contains the model that is used e.g. models/<model name>/inference_graph/frozen_inference_graph.pb}
LABEL_MAP_PATH={Path to label map proto, e.g. data/deepfashion_label_map_fine.pbtxt.}
NUM_CLASSESS={Number of classes the labels have. e.g. 50 FOR DeepFashion}
MIN_SCORE={Threshold FOR minimum score that is considered a True positive. e.g. 0.3}

python inference/Object_detection_webcam.py \
    --ckpt_path ${CKPT_PATH} \
    --label_map_path ${LABEL_MAP_PATH} \
    --num_classes ${NUM_CLASSESS} \
    --min_score ${MIN_SCORE}

Inferencing on an Image

# From the project root directory

CKPT_PATH={Path to frozen detection graph .pb file, Which contains the model that is used e.g. models/<model name>/inference_graph/frozen_inference_graph.pb}
LABEL_MAP_PATH={Path to label map proto, e.g. data/deepfashion_label_map_fine.pbtxt.}
NUM_CLASSESS={Number of classes the labels have. e.g. 50 FOR DeepFashion}
MIN_SCORE={Threshold FOR minimum score that is considered a True positive. e.g. 0.3}
IMAGE_PATH={Path to an image, a png file is used In the paper}

python inference/Object_detection_image.py \
    --ckpt_path ${CKPT_PATH} \
    --label_map_path ${LABEL_MAP_PATH} \
    --num_classes ${NUM_CLASSESS} \
    --min_score ${MIN_SCORE} \
    --image_path ${IMAGE_PATH}

tyrng/deepfashionDetection