Awesome

Grad-CAM implementation in Tensorflow

This repo is a TensorFlow implementation of Gradient class activation maps(Grad-CAM[1]), one of visualization techniques for deep learning networks.

This repo is based on Torch and Keras versions of Grad-CAM.

Requirements

• Python3.x
• Tensorflow 1.x
• tensorflow-vgg (which includes the pretrained(using the Imagenet dataset) VGG16 classification model file VGG16.npy(see the README on how to download it))

Usage

python grad-cam-tf.py <path_to_image> <path_to_VGG16_npy> [top_n]

• path_to_image: an image for which Grad-CAM is calculated.
• path_to_VGG16_npy: path to the pretrained VGG16 model data provided in tensorflow-vgg
• top_n: Optional. Grad-CAM is calculated for each 'top_n' class, which is predicted by VGG16.

The following images are saved to the same directory as path_to_image:

• input image overlaid by Grad-CAM heatmap with the suffix gradcam appended to the path_to_image.
• black-and-white Grad-CAM heatmap(suffix: heatmap)
• segmented image by Grad-CAM heatmap(suffix: segmented)
• Guided Backpropagation[2] (suffix: guided_bprop)
• Also rank of predicted class, the class name and the probability are appended as suffix to each file names above.

Examples

• Example image from the original implementation:

the most probable class: 'boxer'(242)

The below are original image, image overlaid by Grad-CAM heatmap and segmented image by Grad-CAM heatmap.

• Image from [3]:

the most probable class: 'desktop computer'(527)

ground truth(the second probable class): 'desk'(526)

• Image from fatal crash by Uber self-driving SUV:

the most probable class: 'traffic_light'(920)

The below are two predicted classes which are related bicycles.

the 6-th probable class: 'motor_scooter'(670)

the 19-th probable class: 'moped'(665)

Lines below are the top 20 classes predicted VGG16 model. As you can see, VGG16 fails to detect bicycle related objects from this picture(though motor_scooter is detected with low probability).

0: class id: 920, class name: traffic_light, probability: 0.105, synset: n06874185 traffic light, traffic signal, stoplight
1: class id: 627, class name: limousine, probability: 0.092, synset: n03670208 limousine, limo
2: class id: 475, class name: car_mirror, probability: 0.039, synset: n02965783 car mirror
3: class id: 818, class name: spotlight, probability: 0.038, synset: n04286575 spotlight, spot
4: class id: 971, class name: bubble, probability: 0.037, synset: n09229709 bubble
5: class id: 670, class name: motor_scooter, probability: 0.027, synset: n03791053 motor scooter, scooter
6: class id: 656, class name: minivan, probability: 0.024, synset: n03770679 minivan
7: class id: 845, class name: syringe, probability: 0.024, synset: n04376876 syringe
8: class id: 864, class name: tow_truck, probability: 0.022, synset: n04461696 tow truck, tow car, wrecker
9: class id: 867, class name: trailer_truck, probability: 0.018, synset: n04467665 trailer truck, tractor trailer, trucking rig, rig, articulated lorry, semi
10: class id: 555, class name: fire_engine, probability: 0.016, synset: n03345487 fire engine, fire truck
11: class id: 468, class name: cab, probability: 0.016, synset: n02930766 cab, hack, taxi, taxicab
12: class id: 407, class name: ambulance, probability: 0.015, synset: n02701002 ambulance
13: class id: 736, class name: pool_table, probability: 0.013, synset: n03982430 pool table, billiard table, snooker table
14: class id: 836, class name: sunglass, probability: 0.013, synset: n04355933 sunglass
15: class id: 772, class name: safety_pin, probability: 0.010, synset: n04127249 safety pin
16: class id: 837, class name: sunglasses, probability: 0.010, synset: n04356056 sunglasses, dark glasses, shades
17: class id: 754, class name: radio, probability: 0.010, synset: n04041544 radio, wireless
18: class id: 665, class name: moped, probability: 0.010, synset: n03785016 moped
19: class id: 444, class name: bicycle-built-for-two, probability: 0.009, synset: n02835271 bicycle-built-for-two, tandem bicycle, tandem

I think Uber has built in-house classification models for self-driving cars. But detecting persons on roads especially at night could be difficult only by visible light. As reported in Report: Uber's Self-Driving Car Sensors Ignored Cyclist In Fatal Accident it seems a woman crossing the street on a bicycle was ignored.

Detectors using an infrared camera are also needed to avoid such tragic accidents.

Note

For your information, Grad-CAM++, a variant of Grad-CAM, is presented in [4]. The authors provided their implementation. As pointed by Questions for computing of derivatives, their implementation is weird.

References

[1] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, Dhruv Batra. Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization, arXiv, 2016

[2] J. T. Springenberg, A. Dosovitskiy, T. Brox, and M. Riedmiller. Striving for Simplicity: The All Convolutional Net, arXiv, 2014

[3] https://pair-code.github.io/saliency/

[4] Aditya Chattopadhyay, Anirban Sarkar, Prantik Howlader, Vineeth N Balasubramanian. Grad-CAM++: Generalized Gradient-based Visual Explanations for Deep Convolutional Networks, arXiv, 2017