use fcos thresh for batext

Intelligent-Systems-Laboratory · May 11, 2020 · 9223d10 · 9223d10
1 parent 564e7cc
commit 9223d10
Show file tree

Hide file tree

Showing 8 changed files with 11 additions and 11 deletions.
diff --git a/README.md b/README.md
@@ -49,7 +49,7 @@ For more models and information, please refer to MEInst [README.md](configs/MEIn
 
 Name | inf. time | e2e-hmean | det-hmean | download
 ---  |:---------:|:---------:|:---------:|:---:
-[attn_R_50](configs/BAText/TotalText/attn_R_50.yaml) | 11 FPS | 62.7 | 82.8 | [model](https://cloudstor.aarnet.edu.au/plus/s/nyyNRdP7VBYqfgl/download)
+[attn_R_50](configs/BAText/TotalText/attn_R_50.yaml) | 11 FPS | 63.0 | 82.8 | [model](https://cloudstor.aarnet.edu.au/plus/s/nyyNRdP7VBYqfgl/download)
 
 For more models and information, please refer to ABCNet [README.md](configs/BAText/README.md).
 

diff --git a/adet/config/defaults.py b/adet/config/defaults.py
@@ -90,7 +90,6 @@
 _C.MODEL.BATEXT.RECOGNITION_LOSS = "ctc"
 _C.MODEL.BATEXT.RECOGNIZER = "attn"
 _C.MODEL.BATEXT.CANONICAL_SIZE = 96  # largest min_size for level 3 (stride=8)
-_C.MODEL.BATEXT.TEST_CONFIDENCE_THRESHOLD = 0.7 # [0.0 - 1.0]
 
 # ---------------------------------------------------------------------------- #
 # BlendMask Options

diff --git a/adet/evaluation/text_evaluation.py b/adet/evaluation/text_evaluation.py
@@ -53,7 +53,7 @@ def __init__(self, dataset_name, cfg, distributed, output_dir=None):
         elif "ctw1500" in dataset_name:
             self._text_eval_gt_path = "datasets/evaluation/gt_ctw1500.zip"
             self._word_spotting = False
-        self._text_eval_confidence = cfg.MODEL.BATEXT.TEST_CONFIDENCE_THRESHOLD
+        self._text_eval_confidence = cfg.MODEL.FCOS.INFERENCE_TH_TEST
 
     def reset(self):
         self._predictions = []

diff --git a/configs/BAText/CTW1500/Base-CTW1500.yaml b/configs/BAText/CTW1500/Base-CTW1500.yaml
@@ -3,13 +3,10 @@ MODEL:
   BATEXT:
     POOLER_RESOLUTION: (8,128)
     NUM_CHARS: 100
-    TEST_CONFIDENCE_THRESHOLD: 0.65
 DATASETS:
   TRAIN: ("ctw1500_word_train",)
   TEST: ("ctw1500_word_test",)
 INPUT:
   HFLIP: False
-  MIN_SIZE_TRAIN: (640, 672, 704, 736, 768, 800, 832, 864, 896)
-  MAX_SIZE_TRAIN: 1600
   MIN_SIZE_TEST: 800
   MAX_SIZE_TEST: 1024
diff --git a/configs/BAText/CTW1500/attn_R_50.yaml b/configs/BAText/CTW1500/attn_R_50.yaml
@@ -1,6 +1,8 @@
 _BASE_: "Base-CTW1500.yaml"
 MODEL:
   WEIGHTS: "attn_pretrain.pth"
+  FCOS:
+    INFERENCE_TH_TEST: 0.6
   RESNETS:
     DEPTH: 50
   BATEXT:
@@ -13,4 +15,4 @@ SOLVER:
   CHECKPOINT_PERIOD: 10000
 TEST:
   EVAL_PERIOD: 10000
-OUTPUT_DIR: "output/batext/ctw/attn_R_50"
+OUTPUT_DIR: "output/batext/ctw1500/attn_R_50"
diff --git a/configs/BAText/README.md b/configs/BAText/README.md
@@ -8,15 +8,15 @@ Name | inf. time | e2e-hmean | det-hmean | download
 --- |:---:|:---:|:---:|:---:
 [attn_R_50](configs/BAText/CTW1500/attn_R_50.yaml) | 8.7 FPS | 45.2 | 81.3 | [model](https://cloudstor.aarnet.edu.au/plus/s/AexPQx2gB7CrCnQ/download)
 
-* Best E2E result is achieved with threshold 0.6 and Det with 0.5.
+* E2E result is achieved with threshold 0.6 and Det with 0.5.
 
 ### Total Text results with ABCNet.
 
 Name | inf. time | e2e-hmean | det-hmean | download
 ---  |:---------:|:---------:|:---------:|:---:
-[attn_R_50](configs/BAText/TotalText/attn_R_50.yaml) | 11 FPS | 62.7 | 82.8 | [model](https://cloudstor.aarnet.edu.au/plus/s/nyyNRdP7VBYqfgl/download)
+[attn_R_50](configs/BAText/TotalText/attn_R_50.yaml) | 11 FPS | 63.0 | 82.8 | [model](https://cloudstor.aarnet.edu.au/plus/s/nyyNRdP7VBYqfgl/download)
 
-* Best E2E result is achieved with threshold 0.6 and Det with 0.55.
+* E2E result is achieved with threshold 0.6 and Det with 0.55.
 
 
 ## Quick Start 

diff --git a/configs/BAText/TotalText/attn_R_50.yaml b/configs/BAText/TotalText/attn_R_50.yaml
@@ -1,6 +1,8 @@
 _BASE_: "Base-TotalText.yaml"
 MODEL:
   WEIGHTS: "attn_pretrain.pth"
+  FCOS:
+    INFERENCE_TH_TEST: 0.6
   RESNETS:
     DEPTH: 50
   BATEXT:

diff --git a/datasets/README.md b/datasets/README.md
@@ -16,7 +16,7 @@ Run `python prepare_thing_sem_from_lvis.py`, to extract semantic labels from LVI
 ## Text Recognition
 
 - Totaltext training, testing images, and annotations [[link]](https://universityofadelaide.box.com/shared/static/3eq5ti7z45qfq5gu96gg5t1xwh1yrrt7.zip) [[paper]](https://ieeexplore.ieee.org/abstract/document/8270088/) [[code]](https://github.com/cs-chan/Total-Text-Dataset). 
-- CTW1500 training, testing images, and annotations [[link]](https://universityofadelaide.box.com/shared/static/3eq5ti7z45qfq5gu96gg5t1xwh1yrrt7.zip) [[paper]](https://www.sciencedirect.com/science/article/pii/S0031320319300664) [[code]](https://github.com/Yuliang-Liu/Curve-Text-Detector).
+- CTW1500 training, testing images, and annotations [[link]](https://universityofadelaide.box.com/s/yb9red8pi9eszuzqompo593b6zhz87qw) [[paper]](https://www.sciencedirect.com/science/article/pii/S0031320319300664) [[code]](https://github.com/Yuliang-Liu/Curve-Text-Detector).
 - MLT [[images]](https://universityofadelaide.box.com/s/9c4maycxaxo6dd95sfjz087pno3wbnm7)[[annos]](https://universityofadelaide.box.com/shared/static/8hgcrfdvqroqjwy27thu1naez6px82a1.zip) [[paper]](https://ieeexplore.ieee.org/abstract/document/8270168).
 - Syntext-150k (Part1: 54,327 [[imgs]](https://universityofadelaide.box.com/s/1jcvu6z9jojmhzojuqrwxvwxmrlw7uib)[[annos]](https://universityofadelaide.box.com/s/zc73pyzvymqkjg3vkb2ayjol7y5a4fsk).
 - Part2: 94,723 [[imgs]](https://universityofadelaide.box.com/s/ibihmhkzpc1zuh56mxyehad1dv1l73ua)[[annos]](https://universityofadelaide.box.com/s/rk55zheij8ubvwgzg7dfjbxgi27l8xld).)