04-09-2026, 08:07 AM
The corporate code platform GitHub of the well‑known Japanese AI development company ficha.jp contains all of the company’s development code as well as customer‑specific project code.
# **English Translation**
## **1. SDK Core Platforms (Most Critical)**
### **sdk2 ⭐⭐⭐⭐⭐**
**Role:** Second‑generation SDK platform; the company’s core technical framework
**Derived projects:**
- sdk2-desayface — Desay face-recognition version
- sdk2-for-desay — Desay customized version
- sdk2-for-koito — Koito customized version
- sdk2-JVC-adas — JVC ADAS version
- sdk2-JVC-ADAS-Android — JVC Android version
- sdk2-translog-B2C — B2C transaction‑log version
**Features:**
- Multiple customer‑specific branches
- Cross‑platform support (Android, embedded systems)
- Represents the company’s “productized” core
### **SDK4 ⭐⭐⭐⭐⭐**
**Role:** Fourth‑generation SDK platform (latest version)
**Features:** Iterative technical upgrade, likely integrating experience from sdk2
### **AIFramework ⭐⭐⭐⭐⭐**
**Role:** Underlying AI framework; foundation of all AI functionalities
**Status:** Comparable to an “operating‑system‑level” component
---
## **2. ADAS Core Algorithms (Highly Productized)**
### **FichaDet ⭐⭐⭐⭐⭐**
**Role:** Ficha’s proprietary object detector
**Features:**
- Named after the company (Ficha = brand)
- Core intellectual property
- Likely the foundation of many customer projects
### **LaneDet / LaneDetection ⭐⭐⭐⭐⭐**
**Role:** Core lane‑detection algorithm
**Importance:** One of the fundamental ADAS functions
**Usage:** Used in nearly all ADAS customer projects
### **VP_Detection / CalcVp ⭐⭐⭐⭐**
**Role:** Vanishing‑point detection and computation
**Importance:** Key technology for lane and road‑boundary detection
### **TrafficSignLightClassification ⭐⭐⭐⭐**
**Role:** Traffic sign & traffic light classification
**Importance:** Standard ADAS functionality
### **FC_tracker ⭐⭐⭐⭐**
**Role:** Full‑circle tracker
**Features:** Likely a core multi‑object tracking technology
---
## **3. DMS Core Algorithms (Independent Product Line)**
### **DMS Neural Network Series ⭐⭐⭐⭐⭐**
- DMS_InsightFace — Face recognition (based on open‑source InsightFace)
- DMS_BHNet — Behavior detection network
- DMS_BNet — Base network
- DMS_ETNet — Eye‑tracking network
- DMS_EyeNet — Eye‑detection network
- DMS_LMNet — Landmark detection network
**Features:**
- Complete DMS technology stack
- Modular design (face, eyes, behavior separated)
- Indicates DMS is an independent business line
### **FACE-EYE / FACE-EYE2 ⭐⭐⭐⭐⭐**
**Role:** Joint face + eye detection
**Versioning:** FACE‑EYE2 is the improved version
### **GazeEstimation / HeadPoseEstimation ⭐⭐⭐⭐**
**Role:** Gaze estimation + head‑pose estimation
**Importance:** Core functions for driver attention monitoring
### **drowsiness ⭐⭐⭐⭐**
**Role:** Fatigue detection
**Importance:** Critical safety function in DMS
### **dms-demo-library ⭐⭐⭐⭐**
**Role:** DMS demo library (for external demos / customer presentations)
---
## **4. OCR Core Projects (New Business Direction)**
### **pure_ocr_sdk_libtorch ⭐⭐⭐⭐⭐**
**Role:** Pure OCR SDK based on LibTorch
**Features:**
- “Pure” indicates an independent SDK
- LibTorch enables production‑grade deployment
- Represents the latest OCR core
### **table_sdk ⭐⭐⭐⭐**
**Role:** Table‑recognition SDK
**Importance:** Advanced OCR capability (invoices, forms)
### **OCR_API / OCR_SERVER / OCR_WEB ⭐⭐⭐⭐**
**Role:** Three deployment forms of OCR
**Indicates:** Productized with multiple access methods
### **drawing-llm / insurance-llm ⭐⭐⭐⭐**
**Role:**
- drawing-llm: Blueprint / engineering drawing analysis
- insurance-llm: Insurance‑related document analysis
**Features:**
- Introduces LLM technology
- Represents new technical directions
### **mold_design_drawing_analysis ⭐⭐⭐⭐**
**Role:** Mold design drawing analysis
**Indicates:** OCR expansion into industrial applications
---
## **5. Inference Engines (Deployment Core)**
### **ncnn Series ⭐⭐⭐⭐⭐**
- ncnn-classifier — NCNN classifier
- ncnn_int1178 — NCNN integer quantization (likely mixed int8/int11)
- ncnn_tester — NCNN testing tools
**Importance:**
- NCNN is a mainstream inference engine for mobile/embedded
- Quantization (int1178) indicates deep optimization
### **MNN_int1178 ⭐⭐⭐⭐**
**Role:** Alibaba MNN inference engine
**Indicates:** Multi‑engine deployment strategy
### **pytorch2ncnn ⭐⭐⭐⭐**
**Role:** Model conversion tool
**Importance:** Bridge from training to deployment
---
## **6. Major Customer‑Specific Projects**
### **Bosch Series ⭐⭐⭐⭐⭐**
- Bosch-SDK — Bosch SDK (global Tier‑1 giant)
- Bosch-ROS — ROS integration (Bosch requires ROS architecture)
- Bosch-tcpip — TCP/IP communication protocol
- Adrien_Bosch_V4H — V4H platform (Renesas high‑end chip)
**Importance:**
- Bosch = world’s largest Tier‑1 supplier
- Independent SDK indicates large project scale
- V4H = high‑end ADAS solution
### **Clarion Series ⭐⭐⭐⭐**
- Clarion — Main project (Hitachi Group)
- Clarion_RearAEB_H3 — Rear AEB (Automatic Emergency Braking)
**Importance:**
- Hitachi subsidiary; major Japanese automotive supplier
- AEB = safety‑critical function
### **JVC Series ⭐⭐⭐⭐**
- HybridDL_JVC_ADAS_2020 — Hybrid deep‑learning ADAS
- sdk2-JVC-adas — JVC ADAS SDK
**Features:**
- HybridDL likely combines traditional + deep learning
- “2020” indicates long‑term cooperation
### **Toyota Series ⭐⭐⭐⭐**
- ToyotaImageRecog — Toyota image recognition
- ToyotaRoadCrackDetection — Road‑crack detection
**Importance:**
- Toyota = world’s largest automaker
- Road‑crack detection = specialized application
---
## **7. Hardware Platform Adaptation (Technical Barrier)**
### **Jetson Series ⭐⭐⭐⭐⭐**
- jetson-demo — Jetson demo
- adas-dms-jetson-demo-app — Full ADAS + DMS demo
**Importance:**
- NVIDIA Jetson = mainstream automotive AI platform
- Full ADAS+DMS indicates technical maturity
### **Renesas V4H ⭐⭐⭐⭐⭐**
- Adrien_Bosch_V4H — V4H adaptation (Bosch project)
- bosch_communication_sample_v4h
**Importance:**
- V4H = Renesas high‑end ADAS chip
- Widely used by Japanese automakers
- High technical barrier
### **Zynq Platform ⭐⭐⭐**
- koito-fir_in_light-zynq — Koito FIR‑in‑headlamp project
**Specialty:**
- FPGA platform → hardware acceleration capability
- Infrared detection inside headlamp = innovative application
---
## **8. Tools & Frameworks (R&D Efficiency)**
### **AutoBuild ⭐⭐⭐⭐**
**Role:** Automated build system
**Importance:** Infrastructure for multi‑project management
### **FichaCSVtoMSCNN ⭐⭐⭐⭐**
**Role:** Data format conversion (CSV → MSCNN training format)
**Indicates:** Custom data‑annotation workflow
### **data-comparator ⭐⭐⭐**
**Role:** Data comparison tool
**Usage:** Likely for regression testing / accuracy comparison
### **model_repository ⭐⭐⭐⭐**
**Role:** Model repository (centralized management of trained models)
**Importance:** Model version control
---
## **Summary of Core Projects (Top 20)**
### **Platform Level (3)**
1. sdk2 ⭐⭐⭐⭐⭐
2. SDK4 ⭐⭐⭐⭐⭐
3. AIFramework ⭐⭐⭐⭐⭐
### **ADAS Core (5)**
4. FichaDet ⭐⭐⭐⭐⭐
5. LaneDet / LaneDetection ⭐⭐⭐⭐⭐
6. VP_Detection ⭐⭐⭐⭐
7. TrafficSignLightClassification ⭐⭐⭐⭐
8. FC_tracker ⭐⭐⭐⭐
### **DMS Core (4)**
9. DMS_InsightFace ⭐⭐⭐⭐⭐
10. FACE-EYE2 ⭐⭐⭐⭐⭐
11. GazeEstimation ⭐⭐⭐⭐
12. drowsiness ⭐⭐⭐⭐
### **OCR Core (3)**
13. pure_ocr_sdk_libtorch ⭐⭐⭐⭐⭐
14. table_sdk ⭐⭐⭐⭐
15. drawing-llm ⭐⭐⭐⭐
### **Inference Engines (2)**
16. ncnn_int1178 ⭐⭐⭐⭐⭐
17. pytorch2ncnn ⭐⭐⭐⭐
### **Customer Projects (3)**
18. Bosch-SDK ⭐⭐⭐⭐⭐
19. HybridDL_JVC_ADAS_2020 ⭐⭐⭐⭐
20. Clarion_RearAEB_H3 ⭐⭐⭐⭐
---
## **Criteria for Identifying Core Projects**
1. **Naming patterns:**
- Company‑named (FichaDet, FichaCSVtoMSCNN)
- SDK‑based names (sdk2, SDK4, Bosch-SDK)
- Version iterations (FACE-EYE2, SDK4)
2. **Derived branches:**
- sdk2 → 6 customer versions
- DMS → 7 specialized networks
3. **Technical depth:**
- Proprietary algorithms (FichaDet, FC_tracker)
- Quantization optimization (int1178)
- Multi‑platform adaptation (Jetson, V4H, Zynq)
4. **Customer level:**
- Bosch (global Tier‑1 giant)
- Toyota (world’s largest automaker)
- Clarion/Hitachi (major Japanese supplier)
5. **Business completeness:**
- Full ADAS pipeline (detection → tracking → classification)
- Full DMS stack (face → eyes → behavior → fatigue)
- Full OCR workflow (recognition → tables → LLM)
6. **Technical forward‑looking aspects:**
- LLM integration (drawing‑llm, insurance‑llm)
- Hybrid deep learning (HybridDL)
- Multi‑sensor fusion (FIR infrared)
---
## **If Only 10 Projects Could Be Kept**
1. sdk2 / SDK4 — Core platforms
2. AIFramework — AI foundation
3. FichaDet — Proprietary detector
4. DMS_InsightFace — DMS core
5. pure_ocr_sdk_libtorch — OCR core
6. ncnn_int1178 — Inference engine
7. Bosch-SDK — Most important customer project
8. LaneDet — ADAS fundamental
9. GazeEstimation — Key DMS function
10. drawing-llm — Future direction
These represent the company’s technological moat and commercial value.
# **English Translation**
## **1. SDK Core Platforms (Most Critical)**
### **sdk2 ⭐⭐⭐⭐⭐**
**Role:** Second‑generation SDK platform; the company’s core technical framework
**Derived projects:**
- sdk2-desayface — Desay face-recognition version
- sdk2-for-desay — Desay customized version
- sdk2-for-koito — Koito customized version
- sdk2-JVC-adas — JVC ADAS version
- sdk2-JVC-ADAS-Android — JVC Android version
- sdk2-translog-B2C — B2C transaction‑log version
**Features:**
- Multiple customer‑specific branches
- Cross‑platform support (Android, embedded systems)
- Represents the company’s “productized” core
### **SDK4 ⭐⭐⭐⭐⭐**
**Role:** Fourth‑generation SDK platform (latest version)
**Features:** Iterative technical upgrade, likely integrating experience from sdk2
### **AIFramework ⭐⭐⭐⭐⭐**
**Role:** Underlying AI framework; foundation of all AI functionalities
**Status:** Comparable to an “operating‑system‑level” component
---
## **2. ADAS Core Algorithms (Highly Productized)**
### **FichaDet ⭐⭐⭐⭐⭐**
**Role:** Ficha’s proprietary object detector
**Features:**
- Named after the company (Ficha = brand)
- Core intellectual property
- Likely the foundation of many customer projects
### **LaneDet / LaneDetection ⭐⭐⭐⭐⭐**
**Role:** Core lane‑detection algorithm
**Importance:** One of the fundamental ADAS functions
**Usage:** Used in nearly all ADAS customer projects
### **VP_Detection / CalcVp ⭐⭐⭐⭐**
**Role:** Vanishing‑point detection and computation
**Importance:** Key technology for lane and road‑boundary detection
### **TrafficSignLightClassification ⭐⭐⭐⭐**
**Role:** Traffic sign & traffic light classification
**Importance:** Standard ADAS functionality
### **FC_tracker ⭐⭐⭐⭐**
**Role:** Full‑circle tracker
**Features:** Likely a core multi‑object tracking technology
---
## **3. DMS Core Algorithms (Independent Product Line)**
### **DMS Neural Network Series ⭐⭐⭐⭐⭐**
- DMS_InsightFace — Face recognition (based on open‑source InsightFace)
- DMS_BHNet — Behavior detection network
- DMS_BNet — Base network
- DMS_ETNet — Eye‑tracking network
- DMS_EyeNet — Eye‑detection network
- DMS_LMNet — Landmark detection network
**Features:**
- Complete DMS technology stack
- Modular design (face, eyes, behavior separated)
- Indicates DMS is an independent business line
### **FACE-EYE / FACE-EYE2 ⭐⭐⭐⭐⭐**
**Role:** Joint face + eye detection
**Versioning:** FACE‑EYE2 is the improved version
### **GazeEstimation / HeadPoseEstimation ⭐⭐⭐⭐**
**Role:** Gaze estimation + head‑pose estimation
**Importance:** Core functions for driver attention monitoring
### **drowsiness ⭐⭐⭐⭐**
**Role:** Fatigue detection
**Importance:** Critical safety function in DMS
### **dms-demo-library ⭐⭐⭐⭐**
**Role:** DMS demo library (for external demos / customer presentations)
---
## **4. OCR Core Projects (New Business Direction)**
### **pure_ocr_sdk_libtorch ⭐⭐⭐⭐⭐**
**Role:** Pure OCR SDK based on LibTorch
**Features:**
- “Pure” indicates an independent SDK
- LibTorch enables production‑grade deployment
- Represents the latest OCR core
### **table_sdk ⭐⭐⭐⭐**
**Role:** Table‑recognition SDK
**Importance:** Advanced OCR capability (invoices, forms)
### **OCR_API / OCR_SERVER / OCR_WEB ⭐⭐⭐⭐**
**Role:** Three deployment forms of OCR
**Indicates:** Productized with multiple access methods
### **drawing-llm / insurance-llm ⭐⭐⭐⭐**
**Role:**
- drawing-llm: Blueprint / engineering drawing analysis
- insurance-llm: Insurance‑related document analysis
**Features:**
- Introduces LLM technology
- Represents new technical directions
### **mold_design_drawing_analysis ⭐⭐⭐⭐**
**Role:** Mold design drawing analysis
**Indicates:** OCR expansion into industrial applications
---
## **5. Inference Engines (Deployment Core)**
### **ncnn Series ⭐⭐⭐⭐⭐**
- ncnn-classifier — NCNN classifier
- ncnn_int1178 — NCNN integer quantization (likely mixed int8/int11)
- ncnn_tester — NCNN testing tools
**Importance:**
- NCNN is a mainstream inference engine for mobile/embedded
- Quantization (int1178) indicates deep optimization
### **MNN_int1178 ⭐⭐⭐⭐**
**Role:** Alibaba MNN inference engine
**Indicates:** Multi‑engine deployment strategy
### **pytorch2ncnn ⭐⭐⭐⭐**
**Role:** Model conversion tool
**Importance:** Bridge from training to deployment
---
## **6. Major Customer‑Specific Projects**
### **Bosch Series ⭐⭐⭐⭐⭐**
- Bosch-SDK — Bosch SDK (global Tier‑1 giant)
- Bosch-ROS — ROS integration (Bosch requires ROS architecture)
- Bosch-tcpip — TCP/IP communication protocol
- Adrien_Bosch_V4H — V4H platform (Renesas high‑end chip)
**Importance:**
- Bosch = world’s largest Tier‑1 supplier
- Independent SDK indicates large project scale
- V4H = high‑end ADAS solution
### **Clarion Series ⭐⭐⭐⭐**
- Clarion — Main project (Hitachi Group)
- Clarion_RearAEB_H3 — Rear AEB (Automatic Emergency Braking)
**Importance:**
- Hitachi subsidiary; major Japanese automotive supplier
- AEB = safety‑critical function
### **JVC Series ⭐⭐⭐⭐**
- HybridDL_JVC_ADAS_2020 — Hybrid deep‑learning ADAS
- sdk2-JVC-adas — JVC ADAS SDK
**Features:**
- HybridDL likely combines traditional + deep learning
- “2020” indicates long‑term cooperation
### **Toyota Series ⭐⭐⭐⭐**
- ToyotaImageRecog — Toyota image recognition
- ToyotaRoadCrackDetection — Road‑crack detection
**Importance:**
- Toyota = world’s largest automaker
- Road‑crack detection = specialized application
---
## **7. Hardware Platform Adaptation (Technical Barrier)**
### **Jetson Series ⭐⭐⭐⭐⭐**
- jetson-demo — Jetson demo
- adas-dms-jetson-demo-app — Full ADAS + DMS demo
**Importance:**
- NVIDIA Jetson = mainstream automotive AI platform
- Full ADAS+DMS indicates technical maturity
### **Renesas V4H ⭐⭐⭐⭐⭐**
- Adrien_Bosch_V4H — V4H adaptation (Bosch project)
- bosch_communication_sample_v4h
**Importance:**
- V4H = Renesas high‑end ADAS chip
- Widely used by Japanese automakers
- High technical barrier
### **Zynq Platform ⭐⭐⭐**
- koito-fir_in_light-zynq — Koito FIR‑in‑headlamp project
**Specialty:**
- FPGA platform → hardware acceleration capability
- Infrared detection inside headlamp = innovative application
---
## **8. Tools & Frameworks (R&D Efficiency)**
### **AutoBuild ⭐⭐⭐⭐**
**Role:** Automated build system
**Importance:** Infrastructure for multi‑project management
### **FichaCSVtoMSCNN ⭐⭐⭐⭐**
**Role:** Data format conversion (CSV → MSCNN training format)
**Indicates:** Custom data‑annotation workflow
### **data-comparator ⭐⭐⭐**
**Role:** Data comparison tool
**Usage:** Likely for regression testing / accuracy comparison
### **model_repository ⭐⭐⭐⭐**
**Role:** Model repository (centralized management of trained models)
**Importance:** Model version control
---
## **Summary of Core Projects (Top 20)**
### **Platform Level (3)**
1. sdk2 ⭐⭐⭐⭐⭐
2. SDK4 ⭐⭐⭐⭐⭐
3. AIFramework ⭐⭐⭐⭐⭐
### **ADAS Core (5)**
4. FichaDet ⭐⭐⭐⭐⭐
5. LaneDet / LaneDetection ⭐⭐⭐⭐⭐
6. VP_Detection ⭐⭐⭐⭐
7. TrafficSignLightClassification ⭐⭐⭐⭐
8. FC_tracker ⭐⭐⭐⭐
### **DMS Core (4)**
9. DMS_InsightFace ⭐⭐⭐⭐⭐
10. FACE-EYE2 ⭐⭐⭐⭐⭐
11. GazeEstimation ⭐⭐⭐⭐
12. drowsiness ⭐⭐⭐⭐
### **OCR Core (3)**
13. pure_ocr_sdk_libtorch ⭐⭐⭐⭐⭐
14. table_sdk ⭐⭐⭐⭐
15. drawing-llm ⭐⭐⭐⭐
### **Inference Engines (2)**
16. ncnn_int1178 ⭐⭐⭐⭐⭐
17. pytorch2ncnn ⭐⭐⭐⭐
### **Customer Projects (3)**
18. Bosch-SDK ⭐⭐⭐⭐⭐
19. HybridDL_JVC_ADAS_2020 ⭐⭐⭐⭐
20. Clarion_RearAEB_H3 ⭐⭐⭐⭐
---
## **Criteria for Identifying Core Projects**
1. **Naming patterns:**
- Company‑named (FichaDet, FichaCSVtoMSCNN)
- SDK‑based names (sdk2, SDK4, Bosch-SDK)
- Version iterations (FACE-EYE2, SDK4)
2. **Derived branches:**
- sdk2 → 6 customer versions
- DMS → 7 specialized networks
3. **Technical depth:**
- Proprietary algorithms (FichaDet, FC_tracker)
- Quantization optimization (int1178)
- Multi‑platform adaptation (Jetson, V4H, Zynq)
4. **Customer level:**
- Bosch (global Tier‑1 giant)
- Toyota (world’s largest automaker)
- Clarion/Hitachi (major Japanese supplier)
5. **Business completeness:**
- Full ADAS pipeline (detection → tracking → classification)
- Full DMS stack (face → eyes → behavior → fatigue)
- Full OCR workflow (recognition → tables → LLM)
6. **Technical forward‑looking aspects:**
- LLM integration (drawing‑llm, insurance‑llm)
- Hybrid deep learning (HybridDL)
- Multi‑sensor fusion (FIR infrared)
---
## **If Only 10 Projects Could Be Kept**
1. sdk2 / SDK4 — Core platforms
2. AIFramework — AI foundation
3. FichaDet — Proprietary detector
4. DMS_InsightFace — DMS core
5. pure_ocr_sdk_libtorch — OCR core
6. ncnn_int1178 — Inference engine
7. Bosch-SDK — Most important customer project
8. LaneDet — ADAS fundamental
9. GazeEstimation — Key DMS function
10. drawing-llm — Future direction
These represent the company’s technological moat and commercial value.