NSAQA / app.py
laurenok24's picture
Update app.py
553828f verified
raw
history blame contribute delete
No virus
17.9 kB
import os
import io
import pickle
import cv2
import gradio as gr
print(gr.__version__)
from tempSegAndAllErrorsForAllFrames import getAllErrorsAndSegmentation
from models.detectron2.platform_detector_setup import get_platform_detector
from models.pose_estimator.pose_estimator_model_setup import get_pose_estimation
from models.detectron2.diver_detector_setup import get_diver_detector
from models.pose_estimator.pose_estimator_model_setup import get_pose_model
from models.detectron2.splash_detector_setup import get_splash_detector
from scoring_functions import *
from generate_reports import *
from tempSegAndAllErrorsForAllFrames_newVids import getAllErrorsAndSegmentation_newVids, abstractSymbols
from jinja2 import Environment, FileSystemLoader
from PIL import Image, ImageDraw
from io import BytesIO
import base64
# platform_detector = get_platform_detector()
# splash_detector = get_splash_detector()
# diver_detector = get_diver_detector()
# pose_model = get_pose_model()
template_path = 'report_template_tables.html'
dive_data = {}
class CPU_Unpickler(pickle.Unpickler):
def find_class(self, module, name):
if module == 'torch.storage' and name == '_load_from_bytes':
return lambda b: torch.load(io.BytesIO(b), map_location='cpu')
else: return super().find_class(module, name)
dive_data_precomputed = CPU_Unpickler(open('./segmentation_error_data.pkl', 'rb')).load()
# with open('./segmentation_error_data.pkl', 'rb') as f:
# dive_data_precomputed = pickle.load(f)
import sys
import csv
csv.field_size_limit(sys.maxsize)
with open('FineDiving/fine-grained_annotation_aqa.pkl', 'rb') as f:
dive_annotation_data = pickle.load(f)
def extract_frames(video_path):
cap = cv2.VideoCapture(video_path)
# Check if the video file is opened successfully
if not cap.isOpened():
print("Error: Couldn't open video file.")
exit()
# a variable to set how many frames you want to skip
frame_skip = 1
# a variable to keep track of the frame to be saved
frame_count = 0
frames = []
i = 0
while True:
ret, frame = cap.read()
if not ret:
break
if i > frame_skip - 1:
frame_count += 1
# print("frame.shape:", frame.shape)
# resize takes argument (width, height)
frame = cv2.resize(frame, (455, 256))
frames.append(frame)
i = 0
continue
# cv2.imwrite("./tempdata/{}.jpg".format(i), frame)
i += 1
cap.release()
print("frame_count", frame_count)
return frames
def get_key_from_videopath(video):
try:
video_name = video.split('/')[-1]
first_folder = video_name.split('_')[1]
second_folder = video_name.split('_')[2].split('.')[0]
return (first_folder, int(second_folder))
except:
return None
def get_abstracted_symbols_precomputed(video, key, progress=gr.Progress()):
progress(0, desc="Abstracting Symbols")
if video is None:
raise gr.Error("input a video!!")
local_directory = "FineDiving/datasets/FINADiving_MTL_256s/{}/{}/".format(key[0], key[1])
directory = "file:///Users/lokamoto/Comprehensive_AQA/FineDiving/datasets/FINADiving_MTL_256s/{}/{}".format(key[0], key[1])
# dive_data = abstractSymbols(frames, progress=progress, platform_detector=platform_detector, splash_detector=splash_detector, diver_detector=diver_detector, pose_model=pose_model)
# dive_data['frames'] = frames
global dive_data_precomputed
dive_data = dive_data_precomputed[key]
html_intermediate = generate_symbols_report_precomputed("intermediate_steps.html", dive_data, local_directory, progress=progress)
progress(0.95, desc="Abstracting Symbols")
return html_intermediate
def get_abstracted_symbols_calculated(video, progress=gr.Progress()):
progress(0, desc="Abstracting Symbols")
frames = extract_frames(video)
global dive_data
dive_data = abstractSymbols(frames, progress=progress, platform_detector=platform_detector, splash_detector=splash_detector, diver_detector=diver_detector, pose_model=pose_model)
dive_data['frames'] = frames
html_intermediate = generate_symbols_report("intermediate_steps.html", dive_data, frames)
return html_intermediate
def get_abstracted_symbols(video, progress=gr.Progress()):
if video is None:
raise gr.Error("Click on an example diving video first!")
key = get_key_from_videopath(video)
if key is None:
return get_abstracted_symbols_calculated(video, progress=progress)
else:
return get_abstracted_symbols_precomputed(video, key, progress=progress)
def get_score_report_precomputed(video, key, progress=gr.Progress(), diveNum=""):
progress(0, desc="Calculating Dive Errors")
if video is None:
raise gr.Error("input a video!!")
global dive_data_precomputed
dive_data = dive_data_precomputed[key]
local_directory = "FineDiving/datasets/FINADiving_MTL_256s/{}/{}/".format(key[0], key[1])
directory = "file:///Users/lokamoto/Comprehensive_AQA/FineDiving/datasets/FINADiving_MTL_256s/{}/{}".format(key[0], key[1])
intermediate_scores_dict = get_all_report_scores(dive_data)
progress(0.75, desc="Generating Score Report")
print('getting html...')
html = generate_report(template_path, intermediate_scores_dict, directory, local_directory, progress=progress)
progress(0.9, desc="Generating Score Report")
html = (
"<div style='max-width:100%; max-height:360px; overflow:auto'>"
+ html
+ "</div>")
print("returning...")
return html
def get_score_report_calculated(video, progress=gr.Progress(), diveNum=""):
progress(0, desc="Calculating Dive Errors")
global dive_data
frames = extract_frames(video)
dive_data = getAllErrorsAndSegmentation_newVids(frames, dive_data, progress=progress, diveNum=diveNum, platform_detector=platform_detector, splash_detector=splash_detector, diver_detector=diver_detector, pose_model=pose_model)
intermediate_scores_dict = get_all_report_scores(dive_data)
progress(0.75, desc="Generating Score Report")
print('getting html...')
html = generate_report_from_frames(template_path, intermediate_scores_dict, frames)
html = (
"<div style='max-width:100%; max-height:360px; overflow:auto'>"
+ html
+ "</div>")
print("returning...")
progress(8/8, desc="Generating Score Report")
return html
def get_score_report(video, progress=gr.Progress(), diveNum=""):
if video is None:
raise gr.Error("input a video!!")
key = get_key_from_videopath(video)
if key is None:
return get_score_report_calculated(video, progress=progress)
else:
return get_score_report_precomputed(video, key, progress=progress)
def get_html_from_video(video, diveNum=""):
if video is None:
raise gr.Error("input a video!!")
frames = extract_frames(video)
dive_data = abstractSymbols(frames, platform_detector=platform_detector, splash_detector=splash_detector, diver_detector=diver_detector, pose_model=pose_model)
dive_data['frames'] = frames.copy()
html_intermediate = generate_symbols_report("intermediate_steps.html", dive_data, frames)
yield html_intermediate
dive_data = getAllErrorsAndSegmentation_newVids(frames, dive_data, diveNum=diveNum, platform_detector=platform_detector, splash_detector=splash_detector, diver_detector=diver_detector, pose_model=pose_model)
intermediate_scores_dict = get_all_report_scores(dive_data)
print('getting html...')
html = generate_report_from_frames(template_path, intermediate_scores_dict, frames)
html = (
"<div style='max-width:100%; max-height:360px; overflow:auto'>"
+ html_intermediate
+ html
+ "</div>")
print("returning...")
yield html
def get_html_from_finedivingkey(first_folder, second_folder):
board_side = "left" # change!!!
key = (first_folder, int(second_folder))
local_directory = "FineDiving/datasets/FINADiving_MTL_256s/{}/{}".format(key[0], key[1])
directory = "file:///Users/lokamoto/Comprehensive_AQA/FineDiving/datasets/FINADiving_MTL_256s/{}/{}".format(key[0], key[1])
print("key:", key)
diveNum = dive_annotation_data[key][0]
pose_preds, takeoff, twist, som, entry, distance_from_board, position_tightness, feet_apart, over_under_rotation, splash, above_boards, on_boards, som_counts, twist_counts, board_end_coords, diver_boxes = getAllErrorsAndSegmentation(first_folder, second_folder, diveNum, board_side=board_side, platform_detector=platform_detector, splash_detector=splash_detector, diver_detector=diver_detector, pose_model=pose_model)
dive_data['pose_pred'] = pose_preds
dive_data['takeoff'] = takeoff
dive_data['twist'] = twist
dive_data['som'] = som
dive_data['entry'] = entry
dive_data['distance_from_board'] = distance_from_board
dive_data['position_tightness'] = position_tightness
dive_data['feet_apart'] = feet_apart
dive_data['over_under_rotation'] = over_under_rotation
dive_data['splash'] = splash
dive_data['above_boards'] = above_boards
dive_data['on_boards'] = on_boards
dive_data['som_counts'] = som_counts
dive_data['twist_counts'] = twist_counts
dive_data['board_end_coords'] = board_end_coords
dive_data['diver_boxes'] = diver_boxes
dive_data['diveNum'] = diveNum
dive_data['board_side'] = board_side
intermediate_scores_dict = get_all_report_scores(dive_data)
html = generate_report(template_path, intermediate_scores_dict, directory, local_directory)
html = (
"<div style='max-width:100%; max-height:360px; overflow:auto'>"
+ html
+ "</div>")
return html
## gradio where we input a video ###
# with gr.Blocks() as demo_new:
# gr.Markdown(
# """
# # NS-AQA
# This system takes in a diving video, and outputs a detailed report summarizing each component of the dive and how we evaluated it. We first abstract the necessary symbols, and then proceed to score the dive.\n
# Paper: *insert link to paper* \n
# Code: *insert github link*
# """)
# with gr.Row():
# with gr.Column():
# gr.Markdown(
# """
# ## Step 1: Abstract Symbols
# We first abstract the necessary visual elements from the provided diving video. This includes the platform, splash, and the pose estimation of the diver.
# """
# )
# video = gr.Video(label="Video", format="mp4", include_audio=False)
# abstract_symbols_btn = gr.Button("Abstract Symbols", variant='primary')
# symbol_output = gr.HTML(label="Output")
# examples = gr.Examples(examples = [['01_10.mp4'], ['01_11.mp4'], ['01_16.mp4'], ['01_33.mp4'], ['01_140.mp4']], inputs=[video])
# with gr.Row():
# gr.Markdown(
# """
# ## Step 2: Calculate Logic-Based Errors and Generate Detailed Score Report
# """
# )
# get_score_btn = gr.Button("Get Score", interactive=False, variant='secondary')
# score_report = gr.HTML(label="Output")
# # get_score_report_btn = gr.Button("Get Score Report")
# # video.change(fn=enable_get_score_btn, inputs=get_score_btn, outputs=get_score_btn)
# video.change(fn=disable_get_score_btn, inputs=get_score_btn, outputs=get_score_btn)
# video.change(fn=enable_get_score_btn, inputs=abstract_symbols_btn, outputs=abstract_symbols_btn)
# abstract_symbols_btn.click(fn=get_abstracted_symbols, inputs=video, outputs=symbol_output).success(fn=enable_get_score_btn, inputs=get_score_btn, outputs=get_score_btn)
# symbol_output.change(fn=disable_get_score_btn, inputs=abstract_symbols_btn, outputs=abstract_symbols_btn)
# symbol_output.change(fn=enable_get_score_btn, inputs=get_score_btn, outputs=get_score_btn)
# get_score_btn.click(fn=get_score_report, inputs=[video], outputs=score_report)
def enable_get_score_btn(get_score_btn):
return gr.Button(interactive=True, variant="primary")
def disable_get_score_btn(get_score_btn):
return gr.Button(interactive=False, variant="secondary")
#### demo precomputed ########
with gr.Blocks() as demo_precomputed:
gr.Markdown(
"""
# Neuro-Symbolic Olympic Diving Judge
Authors: Lauren Okamoto, Paritosh Parmar \n
This system not only scores an Olympic dive, but outputs a detailed report summarizing each component of the dive and how we evaluated it. We first abstract the necessary symbols, and then proceed to score the dive.\n
See more details on the system by watching this [Video](https://youtu.be/NDtdtguUzjQ). \n
Technical report: [Report](https://arxiv.org/abs/2403.13798). \n
Github: [Code](https://github.com/laurenok24/NSAQA).
""")
gr.Markdown(
"""
## Step 1: Neural Symbol Abstraction
We first abstract the necessary visual elements from the provided diving video. This includes the platform, splash, and the pose estimation of the diver.
"""
)
# with gr.Row():
gr.HTML(
"""
<table>
<tr>
<td>
Platform
<img src='file/platform.png' height='90'>
</td>
<td>
The location of the platform is crucial to determine when the diver leaves the platform, thus starting their dive.
It is also important to assess how close the diver comes to its edge, which is relevant to scoring.
</td>
<td>
Pose Estimation of Diver
<img src='file/pose_estimation.png' height='70'>
</td>
<td>
The pose of the diver in the sequence of video frames is critical to understanding and assessing the dive.
We obtain 2D pose data with locations of various body parts to recognize sub-actions being performed by the diver, such as a somersault, a twist, or an entry, and also assess the quality of that sub-action.
</td>
<td>
Splash
<img src='file/splash.png' height='90'>
</td>
<td>
Splash at entry into the pool is a conspicuous visual feature of a dive.
The size of the splash is an important element in traditional scoring of dives.
A large splash mars the end of a dive and also likely indicates a flaw in form at water entry.
</td>
</tr>
</table>
"""
)
gr.Markdown(
"""
1. Select one of the example diving videos.
2. Hit the **Abstract Symbols** button. The symbols abstracted will appear to the right of the diving video.
"""
)
with gr.Row(variant='panel'):
with gr.Column():
video = gr.Video(label="Video", format="mp4", include_audio=False, sources=["upload"], interactive=False)
examples = gr.Examples(examples = [['01_10.mp4'], ['01_11.mp4'], ['01_16.mp4'], ['01_33.mp4'], ['01_76.mp4'], ['01_140.mp4']], inputs=[video], label="Click on one of the following diving videos")
symbol_output = gr.HTML(label="Output")
abstract_symbols_btn = gr.Button("Abstract Symbols", variant='secondary')
gr.Markdown(
"""
## Step 2: Calculate Rules-Based Errors and Generate Detailed Score Report
Using the abstracted symbols, we calculate different "errors" of the dive.
These errors are: **feet apart; height off board; distance from board; somersault position tightness; knee straightness; twist position straightness; over/under rotation; straightness of body during entry; and splash size.**
Each error is scored on a scale of 0-10, and are then averaged to reach a final score for the dive.
We then programmatically generate a detailed performance report containing different aspects of the dive, their percentile scores, and visual evidence.
It can be helpful for a number of reasons including as a support to human judges and as an educational tool to teach coaches, athletes, and judges how to score.
1. Click the **Generate Score Report** button. The Score Report will be generated below. (Abstract Symbols first if you haven't already!)
"""
)
# with gr.Row():
get_score_btn = gr.Button("Generate Score Report", interactive=False)
score_report = gr.HTML(label="Report")
# get_score_report_btn = gr.Button("Get Score Report")
video.change(fn=disable_get_score_btn, inputs=get_score_btn, outputs=get_score_btn)
video.change(fn=enable_get_score_btn, inputs=abstract_symbols_btn, outputs=abstract_symbols_btn)
abstract_symbols_btn.click(fn=get_abstracted_symbols, inputs=video, outputs=symbol_output).success(fn=enable_get_score_btn, inputs=get_score_btn, outputs=get_score_btn)
symbol_output.change(fn=disable_get_score_btn, inputs=abstract_symbols_btn, outputs=abstract_symbols_btn)
symbol_output.change(fn=enable_get_score_btn, inputs=get_score_btn, outputs=get_score_btn)
get_score_btn.click(fn=get_score_report, inputs=video, outputs=score_report)
############################################################################################################################################
demo_precomputed.queue()
demo_precomputed.launch(share=True, allowed_paths=["."])
######### gradio where we input first and second folder ##
# demo = gr.Interface(
# fn=get_html_from_finedivingkey,
# inputs=["text", "text"],
# outputs=["html"],
# )
# demo.launch(share=True, enable_queue=True,)